<p class="lede">I came to AI late. While most people were figuring out how to cheat on homework with GPT-3.5, I was on a year and a half away from screens. Skating. Travelling. Learning Japanese. Recovering from burnout, in the way you have to recover from burnout, which is by getting away from the thing that broke you in the first place.</p>
<p>By the time I came back, the cheating-on-homework era was over. My first real use of AI was at a top consulting firm, thrown into the deep end with no warm-up.</p>
<p>Consulting at that level is not a place where you get to learn things slowly. Managers are busy. Projects are intense. The only sustainable way to survive is to self-solve, fast, across a range of problems that changes day to day. So I used AI the way you use a survival tool. Not curiosity. Necessity. A new domain on Monday, a new framework on Tuesday, a new analysis on Wednesday. None of them familiar. All of them due that night.</p>
<p>What that produced was an intuition I could not have built any other way. Not from playing. From needing the thing to work, at speed, on real stakes, across an absurd range of use cases.</p>
<h2>The realisation</h2>
<p>The deeper realisation came later. AI worked for me because of how I think, not despite it. The way my brain wants to run, which is wide and fast and restless and jumping between contexts, was exactly what got rewarded by the conditions I was working in. Conventional advice would have called that a problem to manage. The pace of the work made it the only viable mode.</p>
<div class="pull-quote">The MBB world prizes convergence. Pyramid principle. Drive to the answer. That is not wrong. It is incomplete.</div>

<p>Convergence without divergence sounds tidier but it produces a narrower set of possible answers. You converge on the best version of what you already considered. You miss the version you never thought to consider. The frame you never tested. The angle that would have changed the recommendation.</p>
<p>AI restores the missing half. Cheaply, and at speed. For people who could never sit still long enough to systematise their own divergence, it closes a loop that was open their whole career.</p>
<h2>The advice that no longer applies</h2>
<p>The conventional career advice for people whose brains do not sit still is: pick a lane, narrow your focus, learn discipline. That advice was built for a world where divergence was expensive and convergence was the bottleneck. AI flips both. Divergence is cheap now. Capturing it is what is hard. And capturing it is exactly what the restless mind could never do alone.</p>
<p>If you have been told your whole career to slow down and pick a lane, you probably do not need to slow down. You need a system that can keep up with you.</p>
<div class="pull-quote">For people structurally unsuited to "pick a lane and go deep," that system finally exists.</div>

<p>The trait was always the asset. The infrastructure around it was missing. That is the part that is new.</p>


On AI, restless minds, and the half of thinking that consulting forgets

Why being lost is an advantage in the age of AI

<p class="lede">The common assumption is that cheaper production is good for design. Cheaper production is good for production. Those are not the same thing.</p>
<p>Code and design fail differently. A bad line of code is a ticket. Isolate it, fix it, ship the patch. The rest of the product sits untouched. Design sets a direction. Usability. Intent. The felt experience of a product against what a customer was expecting when they opened it. That direction has to be right before you build, not after. Move too far in the wrong direction and the cost is not a tweak. It is a rethink.</p>
<p>AI accelerates the build. It does not pause to ask whether you are building the right thing.</p>
<h3>Design is the surface customers actually experience</h3>
<p>Code is hidden. Design is not. Customers interact with design directly, every session, every screen. The feedback loops — social, word-of-mouth, reviews — are faster than any internal QA cycle. There is nowhere to buffer the consequences of a bad design decision between the decision and the customer.</p>
<p>The speed of that feedback is not an abstraction. In February 2025, Duolingo&#39;s &quot;Duo Death&quot; campaign generated over a billion organic views. By May, after the company announced an AI-first operating model in language that broke with what customers had built a relationship with, it had lost over 400,000 TikTok followers in weeks. The CEO acknowledged on the Q2 2025 earnings call that the controversy had dampened customer growth. Same product. Same engineering. A design and identity decision changed the brand&#39;s relationship with its customers faster than any product change could have repaired it.</p>
<h3>The first thing AI produces is generic</h3>
<p>Dylan Field, co-founder and CEO of Figma, put it plainly in a recent interview:</p>
<blockquote>
<p><em>&quot;The first thing AI gives you is generic by definition because it&#39;s the average of everything it&#39;s seen.&quot;</em></p>
</blockquote>
<p>Generic is fine as a starting point, if the person holding the tool has a point of view strong enough to push beyond it. It is a liability if they do not. AI compresses the distance from brief to output. It does not supply the judgment about whether the output is right.</p>
<p>Field&#39;s frame for this is clay. AI output is raw material. Someone with craft and a genuine perspective has to shape it. Without that shaping, you ship the average. Faster.</p>
<h3>Same customers. Different read on the room.</h3>
<p>In May 2024, Apple released a campaign for the new iPad Pro called &quot;Crush.&quot; A hydraulic press compressed musical instruments, paint, cameras, and creative tools down to the thinness of the new device. The creative community read it as a metaphor for what technology does to human creativity. The backlash was severe enough that Apple cancelled planned TV spots and issued a public apology. Tor Myhren, Apple&#39;s VP of Marketing Communications, acknowledged the ad had <em>&quot;missed the mark.&quot;</em></p>
<p>Samsung responded within days. A musician walks through the wreckage of the crushed tools, picks up a damaged guitar, and begins to play from sheet music displayed on a Galaxy tablet. The tagline: <strong>&quot;Creativity cannot be crushed.&quot;</strong></p>
<p>Same industry. Same customer base, the creative professionals who buy tablets to work. One brand read what those customers were feeling about AI and creative tools in 2024. One didn&#39;t. The design decision, and the instinct behind it, was the entire difference.</p>
<blockquote>
<p>The gap between bad and great designers widens with AI, not narrows. AI amplifies direction. It does not correct it.</p>
</blockquote>
<h3>Bad designers don&#39;t get better. They get louder.</h3>
<p>AI assistance has raised the floor on engineering. Weak engineers produce more passable output than before. The same does not hold for design, because design is not primarily a production problem. What matters is the quality of judgment going in. That is unchanged.</p>
<p>A great designer with AI tools ships a sharper vision faster. A weak designer ships a confident, high-fidelity version of the wrong thing at scale. Coca-Cola&#39;s AI Christmas campaign was widely called <em>&quot;soulless&quot;</em> — departing from the nostalgic, human-driven ads the brand had built its holiday identity around. Valentino&#39;s AI-generated campaign for its DeVain line was panned as <em>&quot;cheap&quot;</em> and <em>&quot;tacky&quot;</em>, the precise opposite of what a heritage luxury brand&#39;s audience expects. Neither was a production failure. Both were direction failures, built quickly.</p>
<p>Jaguar&#39;s 2024 rebrand — abstract visuals, no car in the launch video, heritage stripped out — resulted in 49 vehicles registered across Europe in April 2025. That is not a bug. That is the cost of a first principle set wrong and built on fast.</p>
<h3>The implication for how you invest</h3>
<p>When building was the bottleneck, the logic was: ship something working first, design it better later. When building is cheap, that logic inverts. Moving fast in the wrong design direction is the most expensive bet available. The speed AI provides accelerates the consequence of a bad first principle, not just the velocity of a good one.</p>
<p>Lowering the cost of production is not the same as lowering the cost of a bad decision. For code, those two things have moved together. For design, they have not.</p>
<p>That is the part worth understanding before the next build starts.</p>


Why lowering the cost of production does not lower the cost of a wrong first principle

AI doesn't fix bad design. It scales it.

<p class="lede">The standard model of mentoring runs in one direction. Senior people develop junior people. The assumption underneath it is that seniority and knowledge travel together. They don&#39;t.</p>
<p>Not consistently. Not in the places where it matters most.</p>
<h3>The opportunity cost argument</h3>
<p>A CEO&#39;s job is to make better decisions than the organisation would make without them, set the conditions for good culture, manage strategic tradeoffs, and build the relationships that open doors. The role does not optimise for being the best engineer, designer, marketer, or operator in the building. Most well-run organisations need their senior leaders in the top 50 to 1 percent of competence in their domain. Very few need them in the top 0.1 percent of any specific craft.</p>
<p>That means by definition, in any well-built team, there are people who know more than the most senior leader about things that matter. Not as a failure. As a design feature.</p>
<p>If that is true, and it is, then the people who belong in the room are not only the people you can develop. They include the people who develop you.</p>
<h3>One-way mentoring is an ego structure</h3>
<p>Hiring people you can mentor is a filter that optimises for your comfort, not the organisation&#39;s capability. It selects for people whose current ceiling sits below yours. In stable environments with slow-moving domains, that is manageable. In fast-moving ones, the ceiling gets hit quickly and the cost shows up in the decisions you make without the input you didn&#39;t know you needed.</p>
<p>The best hire is the one who expands the organisation&#39;s knowledge in a direction it didn&#39;t already have. A hire who adds no new capability, no new perspective, no new ground-level signal is a coordination cost with a salary. That is not a recruiting philosophy. It is just the opportunity cost argument, taken seriously.</p>
<h3>The founder dynamic</h3>
<p>There is a structural version of this that makes one-way mentoring not just suboptimal but unavoidable to push against.</p>
<p>Younger founders optimise for ambition and risk tolerance. They are native to the technologies they are building with. They are not carrying the weight of previous cycles, which is both their advantage and their gap. The people they hire to fill those gaps will often be older, more experienced, track-record-bearing. The mentoring relationship is already inverted by the structure of the problem.</p>
<p>Pretending it isn&#39;t is the failure mode.</p>
<p>This is true more broadly than founders. It is true of anyone building a team in a domain that is moving faster than any single career has moved through it. The people who know the most about what is relevant right now are often not the ones with the longest track records.</p>
<blockquote>
<p>The question is not whether you have more experience than the people you hired. It is whether you are learning from them. If you are not, the org chart is doing more work than it should.</p>
</blockquote>
<h3>Hierarchy is a routing system</h3>
<p>Hierarchy is misread as power. It is not. Hierarchy is information architecture. A decision-routing system that ensures the right call gets made at the right level and the right communication reaches the right people. That is its entire function.</p>
<p>The confusion of hierarchy with authority produces a specific and expensive failure mode. Directives go down. Context goes up, filtered and polished at each layer before it arrives. Everyone in the middle optimises for what their manager wants to hear. The actual signal degrades. By the time a real problem reaches someone with the authority to fix it, it has been translated into something manageable.</p>
<p>Two-way mentoring fixes part of this. Leaders who are genuinely learning from their teams receive unfiltered, ground-level signal. The learning is the channel, not just the outcome.</p>
<h3>The exposed archetype</h3>
<p>The exposed archetype is the manager who experiences being taught by a direct report as a loss of status. This is the person whose value is structural, not substantive. Remove the org chart and it becomes unclear what remains.</p>
<p>In flat organisations, or fast-moving ones, or ones where the domain is shifting faster than seniority tracks, that type is the first to become expensive. Not because they are unintelligent. Because their value was always the position, not what they brought to it.</p>
<blockquote>
<p>Design the organisation for information to flow in every direction. Not because it is philosophically tidy. Because it is faster. The team that surfaces problems, ideas, and corrections from every layer makes better decisions than the team waiting for instructions from above.</p>
</blockquote>
<p>Hierarchy is the routing. What flows through it is the thing that actually matters.</p>


On opportunity cost, information architecture, and why one-way mentoring optimises for the wrong thing

Hire people who can teach you

<p class="lede">The fastest-scaling companies of the last two years share a strange signature: tiny teams, founder-written code, founder-led GTM. It is tempting to read this as AI letting weak founders punch above their weight. It is not. The rate-limiting step on a great founder was never talent. It was the team they could afford. AI made the team free.</p>
<p>The technical-business hybrid was already pulling ahead. The founder who could code and sell. The operator who could ship the feature and pitch the round in the same week. They were already taking a disproportionate share of breakout outcomes. AI does not create that trend. It bends the curve up until the hybrid stops being a rising archetype and becomes the default.</p>
<h2>The same shape, three times</h2>
<p>This is not the first time the hybrid has won. It is the third.</p>
<p><strong>The industrial era. 1880 to 1940.</strong> Carnegie. Edison. Ford. Engineers who became capitalists, or capitalists who learned the metallurgy. They understood the steel chemistry and the bond market. They understood the assembly line and the labour contract. The intersection was rare and disproportionately powerful. It was also bottlenecked by the team they needed to scale it. Armies of labour. Foremen. Financiers. The archetype existed but the leverage required tens of thousands of supporting roles to reach.</p>
<p><strong>The software era. 1980 to 2020.</strong> Gates. Jobs. Bezos. Musk. The same shape. Coders with commercial instinct. Operators with technical taste. Each of them paired with a co-founder or built a 50-to-500 person company because the work demanded a coalition. Engineering, design, marketing, sales, ops. The hybrid was at the centre, but the centre was always too narrow. The supporting cast was the limit, and the supporting cast cost time and capital and politics.</p>
<p>The forty years between 1940 and 1980 were when the hybrid was most suppressed. The managerial era. Conglomerates. The professional manager class. Specialisation peaked because organisations had grown to the point where coordination was the binding constraint. The hybrid did not vanish. The hybrid had no slot.</p>
<p><strong>The AI era. 2023 onwards.</strong> Same archetype again. But the supporting cast has collapsed into a model. The engineer is on tap. The marketer is on tap. The designer, the analyst, the customer support agent, the legal reviewer. The hybrid no longer needs the team. The hybrid is the team plus a model.</p>
<p>Across all three eras, the bottleneck was never the person. It was the people they needed around them.</p>
<h2>The coordination tax</h2>
<p>Specialisation was never natural. It was a tax.</p>
<p>The reason a 1000-person organisation runs on lanes is that information cannot move fast enough across a flat structure. So it is split. Engineering is one lane. Product is another. Marketing is a third. The hybrid had to squeeze into a lane to function inside the structure. Be a CTO. Be a CEO. Be a &quot;founder&quot; rather than the actual cognitive shape they were.</p>
<p>The squeeze was always lossy. The person who held both halves of the company in a single thought had to break themselves into pieces to fit on the org chart. Every conversion between lanes lost something. The technical implication of a marketing decision arrived three weeks late, through three handoffs, after the moment had passed.</p>
<p>AI removes the coordination tax. The hybrid no longer hands work between people. They run the loop in their head. They write the brief. They ship the feature. They write the launch copy. They read the data. They iterate. The cognitive shape that was always natural is finally what the work demands.</p>
<p>The shape was right. The world was wrong. The world is catching up.</p>
<div class="pull-quote">AI did not create a new archetype. It removed the team the old archetype always needed.</div>

<h2>The unbundled founding team</h2>
<p>The mechanism is concrete. The 50-person founding team is collapsing into a one-person founding team with a model as the rest.</p>
<p>The hybrid wins at the median, not just the extreme. It used to take a Steve Jobs to operate at the intersection without dilution. The bar was unreasonably high because the supporting infrastructure had to be assembled from scratch every time. Now the bar is conviction and a tolerance for shipping in public. The infrastructure is rented, not built.</p>
<p>The exposed archetype is the pure specialist who never bothered to learn the other side. The deep technical operator who wrote off business as someone else&#39;s job. The deep business operator who wrote off the technical layer as someone else&#39;s job. Both used to be safe inside a team that contained the other. Both are now structurally vulnerable. The hybrid does not need them in the same way the structure used to require.</p>
<h2>Context is what AI cannot give you</h2>
<p>Skills are getting commoditised. AI hands anyone any skill on demand, at increasing depth, on increasing schedule. Coding. Copywriting. Financial modelling. Design. The marginal value of any one skill is collapsing toward zero.</p>
<p>What AI cannot hand over is the cross-domain context that knows which technical decision serves which business outcome in this specific situation, with these specific customers, against this specific competitor, on this specific timeline. That context lives in one head or it does not exist. It is not a skill. It is the connective tissue between skills, and it is built only by holding both sides of the work for years.</p>
<p>The pure technical operator becomes the help. The pure business operator becomes the help. The hybrid becomes the principal. Not because the hybrid is smarter. Because they are the only person on the team carrying the bridge in their head, and the bridge is the entire job.</p>
<p>The trickle becomes the flood. The archetype that was always strong is now the default. The hour is theirs.</p>


AI did not invent a new founder archetype. It removed the team the old archetype always needed.

The hybrid's hour

<div class="memo-hero">
  <img src="/posts/salary-cap-hiring.svg" alt="You can't outbid for talent. Stop trying" loading="eager" />
</div>

<p class="lede">The dominant hiring advice in 2026 is to pay top of market for A players. For most companies that is not a strategy. It is a fantasy. You cannot outbid Google, OpenAI, or a tier-one fund. Trying to do it produces mass hiring at market rate, which is the opposite of what you wanted.</p>
<p>The way to think about hiring is as a pricing problem. Talent is a market. The market mis-prices people in predictable, exploitable ways. Your edge is not paying more. It is seeing what others missed.</p>
<h3>The credit-score problem</h3>
<p>Younger candidates get filtered out for lack of data, the same way someone with no credit history gets rejected by a bank. There is nothing wrong with them. The system passes because the system needs legible signal and they have not been around long enough to generate it. Specific career shapes get filtered out for the same reason. Gaps. Non-linear paths. Unusual combinations. Each pass is a pricing error.</p>
<p>The diagnostic question for every candidate is the same. Why has this person not been processed yet, and is the reason real or a market failure? If the reason is real, move on. If the reason is a data gap, you have just found a barrel the market has not priced.</p>
<h3>Barrels and ammunition</h3>
<p>Keith Rabois has a framing that is worth borrowing here. Most great people are ammunition. They execute well inside a defined scope, they are excellent at what they are asked to do, and they make a company stronger when there is a clear barrel pointing them at the right thing. Barrels are different. They take an idea from inception to shipped, they pull people with them, they hold the whole picture in their head. In a company of two hundred people there are usually somewhere between two and seventeen of them.</p>
<p>This is the unit of measurement most companies are getting wrong. Headcount is not the metric. Barrels are the metric. The number of independent things your company can ship in parallel is bounded by the number of barrels, not the size of the team. Mass hiring fails because it stacks ammunition behind a fixed number of barrels and pays the collaboration tax on top. The burn rate balloons. The output does not.</p>
<blockquote>
<p>Headcount is not the metric. Barrels are the metric.</p>
</blockquote>
<h3>Reference checks are the whole job</h3>
<p>Reference checks are the only mechanism that consistently separates barrels from ammunition. Done properly, they are deep. Twenty conversations, not three. Continuing until you get a negative one, not stopping at three positives in a row. The same person, asked the wrong question, gives the wrong answer. The same person, asked the right question, tells you who this candidate actually is.</p>
<p>The &quot;if you were CEO of this company, what would you do differently?&quot; question filters fast. Ammunition cannot answer it well. Barrels do, because they hold the whole business in their head and have been thinking about it the whole time, regardless of whether their job description asked them to.</p>
<h3>The 30-day test</h3>
<p>The calibration loop sits on the other side of the hire. Ask the team thirty days after the hire whether they would make it again. Ask everyone, not just the manager. The answer at thirty days is almost always the same as the answer at one year and at two years. You can know fast whether you got a barrel.</p>
<p>The implication is uncomfortable. Most hiring processes are designed to be careful at the front and forgiving at the back. The opposite is more efficient. Be willing to hire people the legible market has missed. Then be willing to course-correct fast when the loop tells you the call was wrong.</p>
<h3>The closing problem</h3>
<p>The half most people miss is the close. Once you have found a mispriced barrel, you cannot outbid for them either. The same companies that mis-priced them on the way in are about to re-price them on the way out, and you are still the smaller player. So you do not pitch money. You pitch mission. You pitch the specific shape of the role and why it was built for them. They are not betting on you. They are betting on themselves, in a place where they are specifically needed and where the work will compound faster than anywhere else they could go.</p>
<p>This is the part that does not scale into a process. It is a real conversation about what someone is trying to become and why this role accelerates that. The barrels you want are the ones who are betting on themselves. Pitch them on a place to run. Not a salary band.</p>
<h3>The implication</h3>
<p>The companies that compound talent over a decade are the ones that figured out hiring is not a money game. They built the diagnostic muscle to find mispriced barrels. They built the closing muscle to land them on mission. And they stopped competing on the dimension they could not win.</p>
<p>If you are running a hiring process right now and the strategy reads like a bidding war, you are about to lose a war you could not win in the first place. The good news is the war was the wrong frame. The talent is out there. The market just has not seen it yet.</p>


The hiring problem is a pricing problem

You can't outbid for talent. Stop trying.

<div class="memo-hero">
  <img src="/posts/ceo-relentless-force.svg" alt="The CEO's job is the opposite of morale" loading="eager" />
</div>

<p class="lede">Most leadership advice tells you to lift morale when the team is down and celebrate when the team is up. The opposite is closer to the truth. When the company is winning, the CEO pushes harder. When the company is losing, the CEO becomes a coach. The single role of the CEO is offsetting whatever the team is naturally drifting toward, and complacency in a winning team is more dangerous than fear in a losing one.</p>
<p>This sounds like a contrarian take. It is not. It is what high-performing teams already know about themselves.</p>
<h3>Talented people have their own tempo</h3>
<p>Talented people are like elite athletes. They have an internal tempo. Their morale does not drop when the work is hard. It drops when they are cruising. When the bar in front of them is too low. When nobody is asking more of them than they were already going to give. The motivation problem in a high-performing team is almost never that the team is being pushed too hard. It is that the team is being pushed too little, against a standard that has quietly drifted down with the wins.</p>
<p>This is not a personality observation. It is a structural one. The athlete who needs the harder workout is not unusual. They are the norm in any environment where the people are good. The job of the coach, of the manager, of the CEO, is to be the force that meets the internal rhythm of the team and pulls it up to the company&#39;s actual ambition.</p>
<h3>Why the inversion works</h3>
<p>Winning teams contain the seeds of complacency. The wins feel earned. The standard feels secure. The next sprint feels like it can be a bit easier than the last one because the last one worked. None of that is conscious. All of it is real. Without an active force pulling against the drift, the standard slips by ten percent a quarter, and a year in nobody can remember what the bar used to feel like.</p>
<p>Losing teams contain the seeds of panic. The losses feel personal. The pressure compounds. The team starts cutting corners to stop the bleeding, which produces more losses, which produces more pressure. Without an active force absorbing the panic, the team breaks before it can recover.</p>
<p>The CEO is the only role with the standing and the distance to push back against whichever drift is dominant. Push when the team is drifting toward complacency. Coach when the team is drifting toward panic. The mode flip is not a personality trait. It is the function of the job.</p>
<blockquote>
<p>The single role of the CEO is offsetting whatever the team is naturally drifting toward.</p>
</blockquote>
<h3>The exposed archetype</h3>
<p>The exposed archetype is the leader stuck in one mode. The cheerleader who keeps cheering when the team is cruising. The drill sergeant who keeps pushing when the team is already breaking. Both are reading the situation wrong because they are applying a personality, not a function. Both are doing the job that feels easiest to them, not the job the team needs.</p>
<p>The signal that you are a CEO who can do this well is the willingness to push someone hard precisely because they are winning. Most people cannot. Pushing a winner feels mean. It feels ungrateful. It is exactly the move that keeps the standard from drifting and exactly the move most leaders flinch at. The relentless application of force, kindly and deliberately, against complacency specifically, is what keeps tempo from collapsing.</p>
<h3>The implication</h3>
<p>Hiring talented people is not enough. Talented people will set their own tempo, and their tempo will drift if no one is pulling against it. The CEO&#39;s job is being the force that makes their internal rhythm meet the company&#39;s actual ambition.</p>
<p>If you are leading a team right now and you cannot remember the last time you pushed back on someone for cruising, that is a flag. Not a sign of weakness. A sign that the role you are doing is the one that feels easiest to do, not the one the team needs.</p>


Why winning teams need pressure and losing teams need a coach

The CEO's job is the opposite of morale

<div class="memo-hero">
  <img src="/posts/triad-collapsing.svg" alt="The triad is collapsing" loading="eager" />
</div>

<p class="lede">The product-engineer-designer triad has been the canonical structure of software for fifteen years. PMs decide what to build. Designers decide how it looks and feels. Engineers build it. AI is collapsing this structure faster than the language has caught up, and most teams are still hiring against a shape that is already gone.</p>
<p>The middle of the software process is exactly what AI is best at. The actual building. The hand-off coordination. The translation between specs and mocks and code. None of that is the part that disappears when AI is in the loop. It is the seam where the loop closes. What survives, and what gets harder, is the part the triad was always doing in different vocabularies.</p>
<h3>What the triad was actually doing</h3>
<p>PMs were articulating the value proposition as a roadmap. Designers were articulating it as an experience. Engineers were articulating it as a system. All three were arriving at the same thing — what should exist and why anyone should care — from different angles. The triad existed because no single person could hold all three at the resolution the work demanded. Specialisation was a coordination tax that was cheaper than the alternative.</p>
<p>That is what is changing. AI lowers the cost of holding a second and third angle to the point where the coordination tax of the triad starts to look expensive. The skill that wins is not better PMing or better engineering. It is holding all three angles at once.</p>
<h3>The new archetype</h3>
<p>The new archetype is the technical-business CEO. Someone who can decide what to build, shape how it feels, and ship the code, with AI as a second team that does the rote work in the middle. Ramp&#39;s CEO ships as much code as the rest of his team. That is not a curiosity. It is a leading indicator of where company-building is heading.</p>
<p>This is not about everyone becoming a founder. It is about the shape of the work changing across roles. The best engineers in 2026 have business acumen. The best designers ship working code. The best PMs build things themselves. The premium is on the person who refuses to specialise away from any of the three, because each angle they hold is one fewer hand-off, one fewer translation layer, one less seam where the value proposition gets fuzzy.</p>
<blockquote>
<p>The skill that wins is not better PMing or better engineering. It is holding all three angles at once.</p>
</blockquote>
<h3>The exposed archetype</h3>
<p>The exposed archetype is the specialist who optimised for one angle and never built the others. The PM who never wrote code or designed an interface. The designer who handed off Figma files and never shipped pixels. The engineer who refused to think about users. AI raises the floor under all of them, which means the differential value of pure specialisation collapses. The work they were uniquely good at is now adequately done by anyone with a second angle and a model.</p>
<p>The PM is probably the most exposed. Engineers and designers have a craft AI augments. PMs had a coordination job that AI eats outright. The PM who survives is the one who could have been a CEO all along. The PM who does not is the one who optimised for the meeting, not the build.</p>
<h3>What it means for builders</h3>
<p>If you are early in your career, the highest-leverage move is to build the two angles you do not have yet. If you are an engineer, learn to hold the value proposition. If you are a designer, learn to ship code. If you are a PM, build something yourself, end to end, with AI as the second team. The CV that demonstrates two angles is going to outperform the CV that demonstrates one deeply, because the deep version of one is what AI is already doing.</p>
<p>If you are hiring, stop indexing on deepest-in-one-discipline credentials. Hire for triad-spanning instincts. Build small teams of high-spanning people rather than large teams of deep specialists with coordinators between them.</p>
<h3>The implication</h3>
<p>The companies winning the next decade are not the ones with the best PMs or the best engineers or the best designers. They are the ones with the most people who can hold all three perspectives at once, supervise AI doing the middle work, and keep the value proposition coherent from idea to shipped product.</p>
<p>The triad was a workaround for a coordination cost that AI has made cheap. The shape of the work that replaces it is older and harder. It is just being a builder. It is just knowing what to build and why anyone should care. The job description has not caught up. The work already has.</p>


Why the future builder holds product, design, and engineering at once

The triad is collapsing

<div class="memo-hero">
  <img src="/posts/curiosity-and-agency.svg" alt="Curiosity and agency are the only traits that compound now" loading="eager" />
</div>

<p class="lede">Skill, credentials, and even taste have decay rates. The ground shifts every six months and most of what made someone valuable last year is being absorbed by tools this year. The traits that survive this kind of compression are the ones that do not stop running when the ground moves.</p>
<p>There are two of them, and they only work together. Curiosity. Agency. Curiosity without agency is interesting reading and slow drift. Agency without curiosity is fast motion in a stale direction. The combination is what calibrates your mind, your body, and your worldview at the rate the world is actually changing. The people who hold both, in tandem and at speed, are the ones who will absolutely kill it over the next decade. The people who hold one without the other will be left behind by the ones who hold both.</p>
<h3>What curiosity actually is</h3>
<p>Curiosity is the willingness to stay genuinely interested. Not novelty-chasing. Not the dopamine of new tools. Genuine interest in how things actually work, why they are changing, and what is possible now that was not possible last week. It is the trait that makes someone read the actual paper instead of the summary, try the new tool instead of reading the review, and ask the dumb question in the meeting instead of nodding along.</p>
<p>It is also the trait that takes the longest to fake and the longest to lose. You can fake skills on a CV. You cannot fake the question someone asks, unprompted, about the thing in front of them. Curiosity shows up in behaviour over time, in what someone reads, what they try, and what they ship that nobody asked them to ship.</p>
<h3>What agency actually is</h3>
<p>Agency is the willingness to act on what curiosity surfaces. Build the thing. Try the experiment. Ship the post. Have the conversation. Make the call without permission. Most people who are curious are not agentic. They read, they think, they form opinions, and they do not move. The thinking compounds privately and never escapes into the world.</p>
<p>Agency without curiosity is the more visible failure. Action for the sake of action. Shipping for the sake of shipping. Motion without learning. Most &quot;fast&quot; operators are running this version, which is why their output looks high and their compounding rate is low. They are moving fast in directions that stopped mattering eighteen months ago.</p>
<h3>Why they only work together</h3>
<p>Curiosity and agency operating in tandem is what calibrates a person against the world. You see something, you act on it, the act produces feedback, the feedback updates what you see, and the cycle compresses to whatever the rate of change of the world actually is. This is the mechanism. This is the only mechanism that scales when the ground moves.</p>
<p>Curiosity alone produces an updated worldview that never gets tested. Agency alone produces a stale worldview that gets executed at speed. Neither survives a world that won&#39;t sit still. Together, they produce someone whose view of the world is wired directly to their behaviour, which is wired directly back to their view of the world, in a loop that runs as fast as the loop needs to run.</p>
<blockquote>
<p>Curiosity without agency is slow drift. Agency without curiosity is fast motion in a stale direction.</p>
</blockquote>
<h3>The diagnostic test</h3>
<p>There is a simple pattern at strong companies that surprises people. The most senior person is usually the one using the new tools the most. Not delegating. Not directing. Using. When the CMO is in the prompts every day, the marketing function is healthy. When the CEO is shipping code with AI as a second team, the company is moving. The pattern is not coincidence. It is what curiosity-plus-agency looks like at the top of an org chart.</p>
<p>The pattern holds at every level. The IC who is most curious about the new tooling, and most willing to ship something with it before being asked, is usually the one whose work compounds fastest. The manager who is most curious about how their team&#39;s work is changing, and most willing to redesign the team accordingly, is the one whose team adapts. The trait scales across the org because the underlying dynamic — the world won&#39;t sit still — applies to every role.</p>
<h3>What this looks like as a habit</h3>
<p>The test for whether you have it is concrete. What did you ship this month that nobody asked you to ship? What did you read or try this week that was not on any list? What conversation did you have where you came out with a different view than you went in with? If the answers are nothing, the trait is not there yet. If the answers are interesting, the trait is compounding.</p>
<p>It can be built. Not by accident. The way to build it is the same as the way to use it. Be curious about something specific. Act on it before you are sure. Watch what happens. Update. Repeat. The loop is the point. The loop is what compounds.</p>
<h3>The implication</h3>
<p>If you are choosing who to bet on — to hire, to back, to learn from, to become — optimise for the trait that doesn&#39;t decay. The people who were the smartest in 2022 and stopped updating are not who they were three years ago. The people who stayed curious and stayed acting on it have compounded into people you couldn&#39;t have hired. They are not smarter than they were. They are calibrated against a different world, and the gap is going to keep widening.</p>
<p>The good news is the trait is available to anyone willing to run the loop. The bad news is most people will not, because the loop requires being wrong in public and trying things that do not work. That is the cost of admission. The people who pay it absolutely kill it. The people who do not are reading about the people who did.</p>


On why thinking and doing in tandem is the trait of the rapidly changing world

Curiosity and agency are the only traits that compound now

<div class="memo-hero">
  <img src="/posts/design-engineering-fusing.svg" alt="Code is becoming design. Design is becoming code" loading="eager" />
</div>

<p class="lede">To the average person, design and engineering look like opposite crafts. One is visual, intuitive, taste-led. The other is technical, structured, logic-led. This view is correct about how the two fields were taught and hired ten years ago. It is increasingly wrong about how the work is being done now.</p>
<p>The barriers between the two are dissolving from both directions. Engineers are using AI to ship interfaces without designer hand-offs. Designers are using AI tooling to ship working code without engineer hand-offs. The middle ground that used to need two people and a Figma file is increasingly held by one person, with a model doing the seam.</p>
<h3>The spectrum</h3>
<p>The two crafts sit on a spectrum, not in separate boxes. Design at one end. Engineering at the other. The middle is the part of the work where the two have always met — the part where someone has to translate intent into something that runs.</p>
<pre><code>DESIGN ←──────────────────────────────────────→ ENGINEERING

  Pure visual                                          Pure systems
  Brand, identity                                      Infra, architecture
  Static mocks                                         Backend, APIs
  Typography, taste                                    Performance, scale
  Figma, illustration                                  Servers, databases

                          ↑ the middle ↑
                  prototyping, UI build, interaction,
                  design systems, front-end behaviour
              (this is where the fusion is happening fastest)
</code></pre>
<p>The interesting work has always sat in the middle. Building the actual interface. Wiring up the interaction. Making the design hold together when it meets real data. The middle is where the two crafts have always overlapped, and it is where AI is now lowering the cost of holding both at once.</p>
<h3>The timeline</h3>
<p>The fusion is recent and accelerating. A rough shape of how the middle has been collapsing:</p>
<pre><code>2022 ──── 2023 ──── 2024 ──── 2025 ──── 2026 ──── 2028 (likely)

  ↑         ↑         ↑         ↑         ↑          ↑
  │         │         │         │         │          │
GitHub    ChatGPT     Cursor    v0,     design-     the boundary
Copilot   shifts      makes    Replit   to-code     dissolves
ships     &quot;I can      AI       Agent,   agents      for new
to        write       editing  Figma    standard    builders;
engineers code&quot;       a daily  Make     in IDE;     specialist
          mainstream  habit    early    frontier    titles look
                      for      versions models      vintage
                      builders  ship    can hold
                                        a full
                                        design
                                        system
</code></pre>
<p>The pattern: each step makes it cheaper for one craft to operate in the territory of the other. Engineers got copilots first. Designers got generative tools next. Then the tools that turn descriptions into running code reached enough fidelity that a designer could ship without an engineer in the loop, and an engineer could ship a credible interface without a designer in the loop. The fusion is not a single product. It is a series of bridges getting cheaper to cross until people stop noticing they are crossing them.</p>
<h3>What is real and what is hype</h3>
<p>There are still real differences between the two crafts. Designers hold visual taste at a higher resolution than most engineers. Engineers hold systems thinking and architectural rigour at a higher resolution than most designers. These differences are real and worth preserving. The fusion isn&#39;t about erasing them. It is about lowering the cost of one person holding both, which used to be prohibitive.</p>
<p>What is shifting is the <em>cost</em>, not the underlying difference in talent. A great designer who can now ship code is still a designer. They are just no longer dependent on a hand-off to test their work in the real world. A great engineer who can now design a credible interface is still an engineer. They are just no longer blocked by needing a designer in the loop for every iteration.</p>
<blockquote>
<p>The fusion is not about erasing the difference. It is about lowering the cost of holding both.</p>
</blockquote>
<h3>The exposed archetype</h3>
<p>The exposed archetype is the practitioner who treats their discipline as a moat. The designer who refuses to learn how things ship. The engineer who treats UI as someone else&#39;s problem. Both were defended by the cost of crossing over, and that cost is dropping faster than most of them realise. The moat was always borrowed time. It is now actively eroding.</p>
<p>The interesting people to watch are the ones at the boundary. The engineer who ships their own UI and ships it well. The designer who deploys their own code and writes systems that hold together. They don&#39;t have a job title that fits. The fact that they don&#39;t is a leading indicator that the categories are wrong.</p>
<h3>What the next few years probably look like</h3>
<p>A rough forecast, held loosely:</p>
<ul>
<li><strong>2026–2027.</strong> Design-to-code agents become standard inside design tools. Engineers ship more of the interface work. Specialist hand-offs move up the stack — designers focus on systems and brand, engineers focus on architecture and behaviour, and the middle gets absorbed by tooling.</li>
<li><strong>2027–2028.</strong> The new builder profile starts to dominate at early-stage companies. People who build full products solo or in pairs, holding both sides of the spectrum, with AI as the second team. Job titles lag the work.</li>
<li><strong>2028+.</strong> New entrants stop training as one or the other. Bootcamps and university programs begin merging the curricula. The specialist tracks remain at the frontier ends of the spectrum — pure brand on one side, pure systems on the other — and the middle of both fields converges.</li>
</ul>
<p>This timeline will be wrong in detail and right in shape. The direction is set.</p>
<h3>The implication</h3>
<p>For builders, hirers, and people early in their careers: stop treating these as separate tracks. Future builders won&#39;t be delineated by them. The teams that compound the fastest are the ones with the most people who can hold both at once, and the cost of becoming one of those people has never been lower.</p>
<p>The categories will eventually catch up. The job titles will follow. But the work is already changing, and the people doing the work already know it. The language will get there in three years. The trend is here now.</p>


A spectrum, a timeline, and why future builders won't pick a side

Code is becoming design. Design is becoming code.

<div class="memo-hero">
  <img src="/posts/visceral-learning.svg" alt="The lessons that compound are the ones that hurt" loading="eager" />
</div>

<p class="lede">There are two kinds of learning, and they feel similar in the moment. Information from secondary sources — books, podcasts, advice from people who have done the thing — and visceral experience: failures, hard conversations, moments that left a mark. Both feel like learning when you are doing them. Only one compounds.</p>
<p>This is a more uncomfortable claim than it sounds. It means the things you do most of when you say you are learning are not actually changing how you act. And it means the things that do change how you act are mostly things you would not voluntarily sign up for.</p>
<h3>Two kinds of learning</h3>
<p>Information from secondary sources stacks like data. You can recall it on demand. You can repeat it in conversation. You can apply it deliberately when someone asks. What it does not do is change how you act under pressure, when no one is asking. The lessons sit in the front of your mind, available to be summoned when it is convenient. They never make it to the back, where the actual decisions get made.</p>
<p>Visceral learning is different. It comes from experiences emotive enough that you cannot forget them. A failure that scarred. A conversation that landed in your gut. A moment where the world rearranged itself and you came out the other side seeing differently. These leave imprints that operate below conscious recall. They change what you do, not just what you know.</p>
<p>The gap between the two is invisible in the short run. Both feel productive. Both feel like learning. The gap shows up two years later, when one person has changed and the other person has accumulated a lot of notes.</p>
<h3>Failing fast does not compound. Failing hard does.</h3>
<p>The LinkedIn-thread version of &quot;fail fast and learn&quot; misses this entirely. The instinct is right. The mechanism is wrong. Failing fast does not compound unless the failure is deep enough to leave a mark. Most &quot;failures&quot; in modern professional life are too cheap, too consequence-free, too quickly absorbed into a tidy lesson, to actually shift anything.</p>
<p>A pivot that did not work, processed in a retro the next morning, is not a visceral lesson. It is a data point. The visceral lesson is the one that cost you something you cannot easily replace. Time. Reputation. A relationship. A piece of self-image. The cost is what writes the lesson into the back of your mind.</p>
<blockquote>
<p>The lessons that compound are the ones you cannot afford to forget.</p>
</blockquote>
<h3>The exposed archetype</h3>
<p>The exposed archetype is the person who reads constantly, listens to every podcast, takes detailed notes, and never quite changes. They are stacking information at a rate the brain cannot convert into wiring. The volume looks impressive. The compounding rate is near zero. Over five years the gap between them and someone who lived through one hard year becomes uncrossable.</p>
<p>This is not an argument against reading. It is an argument against confusing reading with learning. The reading is the prep. The visceral experience is the test. Without the test, the prep does not get encoded.</p>
<h3>The logical inversion</h3>
<p>If only visceral experience compounds, the implication is uncomfortable. You cannot optimise for low-cost lessons. You have to actively place yourself in environments where the experiences are dense, the stakes are real, and the consequences land. By definition, comfortable environments do not produce visceral lessons. Comfort is the enemy of compounding.</p>
<p>The instincts that protect you from visceral experience — caution, hedging, staying in roles you have outgrown, avoiding the conversation, picking the safer bet — feel safe. They are. They are also a slow erosion of your capacity to learn, because nothing that happens in their shadow leaves a mark.</p>
<p>This does not mean throwing yourself into chaos. It means recognising that the things people instinctively avoid are precisely the experiences that produce the lessons that compound. Public failure. Hard conversations. Decisions you cannot unmake. Moments where you might be wrong in front of people. Each of these is a candidate for a visceral lesson. Each of these is something most career advice quietly tells you to minimise.</p>
<h3>The implication</h3>
<p>If you want to compound, design for the density of meaningful experience, not for the comfort of the path. Pick roles where the stakes are real. Put yourself in conversations you would rather avoid. Take swings where the cost of being wrong is high enough to leave a mark.</p>
<p>The people who compound the fastest are not the ones who read the most. They are the ones who put themselves into situations where the pressure was sustained, the consequences were real, and they came out the other side different. That is what learning, in any sense that compounds, actually looks like. The rest is preparation for it.</p>


Why most “learning” never makes it past the front of your mind

The lessons that compound are the ones that hurt

<div class="memo-hero">
  <img src="/posts/great-pivot.svg" alt="The great pivot" loading="eager" />
</div>

<p class="lede">The most ambitious people in consulting, law, and banking are leaving. Not at the median. At the top. The people who would have been partners in five years are walking out for startups, AI companies, and operating roles. The firms are losing the top decile of their cohort and replacing them with people who would have been second-tier hires five years ago.</p>
<p>The trend is not yet legible inside the firms. Promotion structures still work. Recruiting numbers still hit. The partner track still fills. The metric that matters — the quality of the people the firm is <em>retaining</em> — is the one nobody is measuring, because measuring it would be embarrassing.</p>
<h3>Three things flipped at once</h3>
<p>The marginal value of optionality has reversed. Five years ago, joining BCG or Goldman or Cravath was the high-expected-value bet because the brand bought you a decade of optionality you could deploy whenever you wanted. Startups and operating roles were the high-variance side of the trade. That has flipped. Operating roles have become more legible to recruiters, more respected by peers, and more compounding for the person doing them. The brand premium has not disappeared, but it has compressed, and the discount on the alternative has compressed faster.</p>
<p>AI is commoditising the analytical work faster at the big three than people inside realise. The work that used to take a second-year associate three weeks now takes a partner with the right tools an afternoon. This is not yet showing up in fees, because the firms are still pricing on the old model. It is showing up in the <em>experience</em> of being a junior in those firms, where the work that was supposed to teach you something is increasingly being done by something else, and the part you do is the part that has not yet been figured out by the model.</p>
<p>The prestige tax has stopped pricing in. The salary discount you accepted in exchange for brand on your CV used to buy you something. A door that opened in five years. An interview at a fund. A board seat that came with the firm name. These doors still open, but they open for fewer people, and the people walking through them increasingly have operator credentials rather than firm credentials. The brand still helps. It does not buy what it used to.</p>
<h3>The firms haven&#39;t noticed</h3>
<p>The exposed archetype here is the firm leadership that thinks the talent problem is a compensation problem. They keep raising salaries to retain. The people they are trying to keep are not leaving for money. They are leaving because the work feels like it is compounding more slowly than the work happening outside the firm, and no salary can fix the shape of the work.</p>
<p>This is the conversation the firms are not having with themselves. The retention metric they care about is partner-track attrition, which still looks fine on average. The retention metric that matters is the attrition of the top quartile, which is climbing and which they are explicitly not breaking out, because doing so would force the question of what is wrong with the work itself.</p>
<blockquote>
<p>The firms are not losing the talent war on price. They are losing it on the shape of the work.</p>
</blockquote>
<h3>The harder problem is for the people inside</h3>
<p>The conventional advice to ambitious people inside the firms — stay, build the brand, take the partner shot — was calibrated to a world where the brand still bought you something durable. That world is fading. Not gone. Fading. The new question is whether the firm is still teaching you something the market values, or whether you are paying a tax on optionality you used to need and do not anymore.</p>
<p>The friends I have asking this question — and there are many — share a pattern. They came in for the rigour. They learned the rigour. The rigour was real and useful. And they noticed, somewhere around year three, that the rigour was no longer the thing that was teaching them. The next two years would be the same rigour applied to the same kinds of problems for incremental compensation. Outside the firm, the same two years would be radically different, with radically more learning per unit of time.</p>
<p>I am one of these people. I came up through BCG. The training was real. The pace was real. The thing I noticed by the end of my time there was that I had absorbed what the firm could teach me at the rate it was capable of teaching, and the next year was going to teach me less than the year before. That was not a failure of the firm. It was a structural feature of how big firms compound a junior. The first year is steep. The second year is steep. The third year is less steep. The decision was whether to keep climbing the curve I was on, or to step onto a steeper one.</p>
<h3>What the firms still get right</h3>
<p>What each of the big three trains you for is not nothing. Consulting trains structure, communication, and the ability to operate in ambiguity. Law trains rigour, written precision, and the discipline of seeing risk where others do not. Banking trains numerical fluency and a high-stakes operating tempo. Each of these is real, transferable, and one of the reasons that some of the best founders come from these backgrounds. The firms aren&#39;t dying. The top of the cohort is just leaving earlier.</p>
<p>The interesting question is not whether the firms are still useful as training grounds. They are. The question is how long, and for what next chapter. The shape of the answer that is emerging is two to three years, mid-twenties, then leave. Long enough to absorb the rigour. Short enough that you are out before the curve flattens and before the optionality stops compounding.</p>
<h3>The implication</h3>
<p>For the firms, the implication is uncomfortable. If the top of the cohort is leaving in year three, the partnership pipeline in year ten is built from the people who stayed because they had nowhere better to go. That is not a recipe for the firm of the next decade. The fix requires changing the shape of the work, not the price of it, and changing the shape of the work is much harder than raising salaries.</p>
<p>For ambitious people inside, the implication is sharper. The question is no longer whether to leave. It is when, and into what. The cost of staying too long is no longer zero, the way it used to be, because the world the firm is preparing you for is being rebuilt by the people who left.</p>
<p>The great pivot is not just a wave of resignations. It is a re-pricing of where the most compounding work actually lives. The firms still matter. They just don&#39;t matter the way they used to. The people who have figured this out have already left, or are about to. The firms will eventually notice. By the time they do, the cohort that would have replaced the leavers will already be running the companies that the firms used to advise.</p>


Why the best people are leaving consulting, law, and banking — and what the firms still don't see

The great pivot

<p class="lede">Every six months the generalist vs specialist debate gets relitigated. It is the wrong debate. Businesses have always needed a mix of both, calibrated to use case. AI does not change that.</p>
<p>What AI changes is the bar. The required depth in any specialism. The required breadth in any generalism. Both are higher than they were two years ago, and both are still rising.</p>
<p>Picture the standard 2x2. Generalist breadth on one axis, specialist depth on the other. AI pushes both axes outward. The same job title now demands more of both dimensions to remain economically valuable.</p>
<h2>What this looks like in practice</h2>
<p>Run it through specific roles and the shift gets concrete:</p>
<ul>
<li><strong>Performance marketers.</strong> The bar moved from running paid media well, to running paid media and reading CRO data and directing creative and building measurement.</li>
<li><strong>Engineers.</strong> The bar moved from shipping features, to shipping features and architecting with AI assistance and reviewing AI-generated code with judgment.</li>
<li><strong>Designers.</strong> The bar moved from producing good output, to producing good output and directing AI tooling and holding taste at scale.</li>
</ul>
<p>The shallow generalist was always fragile. They are now actively exposed, because AI is better at shallow breadth than humans.</p>
<p>The pure specialist is fine, but only if their depth has risen with the bar. Specialists running on 2022 depth are losing leverage they do not realise they had.</p>
<h2>Same destination, different routes</h2>
<p>Both lanes converge on the same outcome: clearing the new bar. How you get there is shaped by your personality and your starting point. Generalists by nature should use AI to deepen faster than they could before. Specialists by nature should use AI to broaden faster than they could before. The path differs. The bar does not.</p>
<div class="pull-quote">Forcing yourself into an archetype you are not is the third failure mode. The natural generalist trying to become a specialist for credibility. The natural specialist trying to become a generalist for optionality. Both lose the speed advantage their actual cognition gave them.</div>

<p>The dumbest move is the one most people are making. Declaring the other side dead. Calling yourself the surviving type. And not actually working on either dimension. Generalists telling themselves specialists are obsolete. Specialists telling themselves generalists are shallow. Both staying exactly where they were in 2022.</p>
<h2>The actual work</h2>
<p>Pick your starting lane based on how you actually think, not how you wish you thought. Use AI to clear the new bar in that lane faster than you could before. The lane is real. The lane war is fake. And the bar is moving whether you are ready for it or not.</p>
<p>If you are hiring against a job spec written in 2022, you are going to over-hire on credentials and under-hire on leverage. The slope of the curve is what changed. Most specs have not caught up.</p>


The generalist vs specialist argument is dead. Both sides are missing what actually changed.

Stop the lane wars

<div class="memo-hero">
  <img src="/posts/eats-the-middle.svg" alt="What still compounds when AI eats the middle" loading="eager" />
</div>

<p class="lede">AI is compressing the cost of execution across every value chain. Building software. Producing creative. Drafting documents. Writing code. Generating analysis. The middle of every business — the actual making of the thing — is getting cheaper, faster, and more commoditised by the month.</p>
<p>When the middle commoditises, value pools at the ends. Two moats survive. The why, on one end. Distribution, on the other. Everything in between is borrowed time. Most companies are still investing in the middle because the middle is where the org chart sits, and they have not yet noticed that the part of the org chart they keep adding to is the part AI is busy hollowing out.</p>
<h3>The value chain, drawn</h3>
<p>Every business runs a value chain that turns an idea into something a customer pays for. A simplified version:</p>
<pre><code>   THE WHY              THE MIDDLE              DISTRIBUTION
   ────────             ──────────              ────────────
   judgment             execution               channel
   taste                building                relationships
   articulation         shipping                trust with audience
   vision               operations              demand capture

   ┌─────────┐  ┌──────────────────────┐  ┌─────────────┐
   │         │  │                      │  │             │
   │  WHY    │──│      MIDDLE          │──│DISTRIBUTION │
   │         │  │                      │  │             │
   └─────────┘  └──────────────────────┘  └─────────────┘
       ↑                  ↓                       ↑
   compounds         commoditising              compounds
       │              fast (AI)                    │
       │                                           │
       └─────── value pools at the ends ───────────┘
</code></pre>
<p>The &quot;why&quot; end is the part that decides what should exist and articulates why anyone should care. Taste, judgment, point of view, the resonance of a value proposition. Design and marketing in the broad sense — the disciplines that turn a thing into something people understand and want.</p>
<p>The &quot;distribution&quot; end is the path from a finished thing to a person who pays for it. Channel ownership. Audience trust. Sales motion. The relationships and habits that put the thing in front of the person at the moment they are ready to buy.</p>
<p>The middle is everything between the two. The actual making. Writing the code. Producing the asset. Running the operation. Translating the spec into the artifact.</p>
<h3>Why the middle is on fire</h3>
<p>AI is compressing the cost of the middle to near zero, asymmetrically and unevenly, but undeniably. Code generation. Content generation. Operational tooling. Customer service. Internal analytics. Each of these is a job that used to require a layer of people, and increasingly requires one person and a model. The cost curve on building things is collapsing faster than any team is keeping up with.</p>
<p>This does not mean the middle disappears. It means the middle stops being a moat. A capability that everyone has at near-zero marginal cost is, by definition, not a competitive advantage. It is table stakes.</p>
<blockquote>
<p>A capability that everyone has at near-zero marginal cost is not a moat. It is table stakes.</p>
</blockquote>
<h3>Why the why end compounds</h3>
<p>The why end does not commoditise because it requires a point of view. AI reverts to the mean. Articulating a value proposition that resonates, spotting what should exist before anyone else does, holding the taste to know what is worth keeping — these are the things AI raises the floor on but cannot substitute for. They depend on judgment that has been calibrated against specific customers, specific markets, specific moments in time, and that calibration is what makes a why durable.</p>
<p>The companies with the strongest &quot;why&quot; position are the ones where someone — usually a founder, sometimes a designer or marketer with real authority — has spent enough time inside the problem to hold a sharper view of it than the market does. That sharpness compounds. The next decision they make is informed by the last hundred. AI does not have those hundred. It has averages.</p>
<h3>Why distribution compounds</h3>
<p>Distribution does not commoditise because it depends on relationships, trust, and channel ownership that accumulate over years. You can prompt your way to a landing page in an hour. You cannot prompt your way to an audience that trusts you, a sales motion that works, or a channel where your customers are already paying attention.</p>
<p>The companies with the strongest distribution position are the ones that started building it early and kept investing when it looked unsexy. Newsletter audiences built over a decade. Sales relationships maintained through quarters where they did not pay off. Communities where the trust is genuine because the company has been in them long enough to have history. These positions compound at a rate set by the relationship, not by the model. AI raises the cost of competing on this dimension. It does not let new entrants skip it.</p>
<h3>The exposed archetype</h3>
<p>The exposed archetype is the company whose entire competitive position is &quot;we build it better.&quot; For most companies that built their identity around execution quality, the moat is dissolving in real time and the leadership team has not noticed. Build quality is converging across the industry as AI raises the floor. The company that was three years ahead on execution is now six months ahead, and the gap is closing every quarter.</p>
<p>The middle-of-the-chain investments — the ones that look like efficiency, productivity, and operational excellence — are the ones AI is eating. Doubling down on them is doubling down on a position that is depreciating.</p>
<h3>The new investment shape</h3>
<p>The companies winning the next decade are running a different shape of investment. Heavy at the why end. Heavy at the distribution end. Light in the middle, with AI absorbing the work that used to require a layer of people.</p>
<pre><code>   investment shape, 2018:        investment shape, 2026:

   why ▓▓                         why ▓▓▓▓▓▓▓
   middle ▓▓▓▓▓▓▓▓▓▓              middle ▓▓
   distribution ▓▓▓▓               distribution ▓▓▓▓▓▓▓
</code></pre>
<p>Most companies are still running the 2018 shape. The cost of running it is rising, because the middle is the part AI is making cheap, and a heavy investment in a part that is getting cheap looks like efficiency on a spreadsheet and like a slow erosion of moat in reality.</p>
<h3>The diagnostic</h3>
<p>The diagnostic for any business right now is to trace the chain from &quot;why this exists&quot; to &quot;how this reaches the user&quot; and ask where the compounding is actually happening. If the answer is at the why end (sharper taste, sharper judgment, sharper articulation than the market) or at the distribution end (channels and relationships that are getting harder to replicate), the moat is real. If the answer is somewhere in the middle — better internal tooling, faster execution, more efficient ops — the moat is decaying faster than the leadership team realises.</p>
<p>The companies that figured this out have already moved their investment to the bookends. The ones that have not are still hiring to make the middle better. The first group is going to look enormously more durable in three years. The second group is going to wonder what happened.</p>


The two moats that survive when execution is free

What still compounds when AI eats the middle

<div class="memo-hero">
  <img src="/posts/great-teams.svg" alt="Most great teams aren't made of great people" loading="eager" />
</div>

<p class="lede">The premise most companies start with is wrong. They believe the path to a high-performing team is to hire enough top-top-top operators that the team is composed of nothing else. This is unrealistic at any scale beyond fifteen people, in any market with a salary cap, against any competition with money. The math does not work. The teams that win are not the ones that hired only A-players. They are the ones that figured out what to do with the team they could actually build.</p>
<p>There is a separate piece of this argument about how to find and close mispriced barrels — the rare operators the market has not yet seen. Assume you have done that work. Assume you have a few of them. You still do not have enough. The next question is the one almost no one asks honestly. What kind of environment turns the rest of the team — the good but not extraordinary, the 70th to 80th percentile — into something that produces 95th-percentile output? That is the memo most companies need and almost no one writes.</p>
<h3>The trap on the other side</h3>
<p>Most culture advice is about preventing the bad. Stopping things from breaking. Making sure mistakes do not cascade. Keeping average performers from dragging the team down. This is real and necessary and not the point. The point is that good people have a natural ceiling that is well below great. Most of them will not reach great alone, no matter how motivated they are. The operating model that wins is the one that closes the gap between their natural ceiling and the standard the company actually needs. Lift good to great. Not floor to ceiling.</p>
<p>This is a different argument from the standard one about psychological safety. The standard argument is about making sure good people do not become worse. The argument here is about making sure good people become better than they would be anywhere else. Different problem. Different mechanism.</p>
<blockquote>
<p>The companies that win are not the ones that hired only A-players. They are the ones that figured out what to do with the team they could actually build.</p>
</blockquote>
<h3>The ratio is the constraint</h3>
<p>You need a critical mass of extraordinary talent for any of this to work. Below the threshold, the extraordinary people get pulled down to the team&#39;s level and quit. Above it, the good people get pulled up to the extraordinary people&#39;s level and compound. The exact ratio is context-dependent, but the existence of a threshold is not.</p>
<p>Most teams underestimate it. Two extraordinary people in a team of fifteen is not enough. Five is closer. The math of human attention sets the floor — there is only so much exposure each barrel can give to the people around them, only so many decisions they can shape, only so many moments where their patterns become visible. Below the threshold, the dilution is too high and the standard does not transfer. Above it, the standard becomes the air the team breathes.</p>
<p>The first hiring decision that matters in any growing team is not &quot;who do we hire next.&quot; It is &quot;are we above or below the ratio.&quot; If you are below it, the next hire has to be a barrel. Compromising on this is how teams plateau and stop knowing why.</p>
<h3>Mechanism one — autonomy</h3>
<p>Good people perform at great levels when given room to own things end to end, with the cost of being wrong bounded but real. The trap is the over-controlled environment that prevents learning by preventing failure. Good people in over-controlled environments stay good. They never become great because the loop that produces greatness — try, fail, learn, try again at higher stakes — is broken by the safety net that was supposed to protect them.</p>
<p>Autonomy is not the same as delegation. Delegation is &quot;do this thing.&quot; Autonomy is &quot;this is yours, including the strategic call about whether to do the thing at all.&quot; The difference is whether the person owns the outcome or owns the task. Outcome ownership is what produces the visceral feedback that compounds into judgment. Task ownership produces practice without learning.</p>
<h3>Mechanism two — adjacency</h3>
<p>Good people working closely with great people compound up at a rate they cannot achieve alone. This is not training. It is osmosis. The patterns of how a great operator thinks, decides, and pushes back transfer through proximity and repetition. The good person watches the great person trade off speed against rigour in a real call, and absorbs the trade-off in a way no framework would have taught them.</p>
<p>Adjacency only works when the team is small enough that the great people are visible to the good people every day, and when the work is collaborative enough that the patterns are observable in action, not in retrospective summary. Hide your barrels in their own pod and the rest of the team learns nothing from them. Embed them. Make their work legible. The team levels up around them, slowly at first, and then all at once.</p>
<h3>Mechanism three — strategic context</h3>
<p>The leader gives the team more context than feels comfortable. The why, the bets, the trade-offs, the long view. The things normally held at the executive level because sharing them feels risky. Information asymmetry is what keeps good people from operating at great levels. They cannot make great calls because they do not have the information that would make the great call obvious. Flood them with strategic context, including the parts that feel risky to share, and the operational calls they make get sharper.</p>
<p>This is the move that most leaders flinch at. Sharing the strategy feels like giving away leverage. It is not. The leverage was never the information. The leverage was the judgment that came from holding the information. Share the information and watch the team develop the judgment. The cost of doing this is feeling exposed. The cost of not doing it is a team that executes well on the wrong things, because they were never told what the right things were.</p>
<h3>Safe to fail is three things, layered</h3>
<p>Psychological safety to speak up and admit mistakes is the table-stakes version. Most teams say they have it. Most teams are wrong about that, but the version most companies aspire to is real and worth building.</p>
<p>Operational safety to fail without cascade is the version most teams underinvest in. Bad bets killed fast. Costs of being wrong bounded. Single failures contained so they do not take the whole quarter with them. This is not about culture, it is about systems — the kill criteria that exist before the experiment starts, the budgets ringfenced before the bet is placed, the explicit naming of what would make us shut this down.</p>
<p>Cultural safety to be <em>ambitious</em> is the version almost no one builds. The downside of missing an ambitious shot is bounded. The upside of taking it is celebrated. The worst thing you can do is play it safe, and the second worst is fail at something timid. The third version is what makes good people swing at great. The first two prevent collapse. The third produces compounding.</p>
<blockquote>
<p>The worst thing you can do is play it safe. The second worst is fail at something timid.</p>
</blockquote>
<h3>The exposed archetype</h3>
<p>The exposed archetype is the founder who hires for safety and then complains the team is not ambitious enough. They built an environment where the rational move is to play it safe. They wonder why the team plays it safe. They blame the people. The people are responding correctly to the incentives the environment creates. The fix is not better people. The fix is the environment.</p>
<p>The same archetype shows up in companies that have done the hiring work but skipped the operating-model work. They have their barrels. They have the right ratio. The barrels are getting frustrated and starting to leave, because nothing the company has built around them is using them as a lever. The barrels were supposed to lift the team. The team did not get lifted, because the autonomy is not there, the adjacency is not there, the context is not there, and the safety to be ambitious is not there. The barrels carry the company on their backs for a while and then leave for a place that knows how to use them.</p>
<h3>The implication</h3>
<p>You probably cannot hire your way to great. You can build your way there. The compounding happens when the right ratio of extraordinary talent runs in close proximity to good talent, with autonomy to own outcomes, strategic context to make great calls, and a culture that makes ambitious shots safer than timid ones. None of these is a hiring move. All of them are operating-model moves. They are available to anyone willing to build them.</p>
<p>The hiring loop is half the work. The operating model is the other half. Most companies that fail at scale are failing on the second half, not the first. They have the people. They do not have the environment. The team plateaus at the average of its hires, and the founders blame the talent market when the answer was always sitting on their side of the table.</p>
<p>Build the environment. The team you can actually hire becomes the team that wins.</p>


Why the operating model matters more than the bench

Most great teams aren't made of great people

<p class="lede">The Claude-style design loop is now available to anyone with a browser. Describe what you want. AI generates working code. See it live. Describe what to change. It iterates. The loop that used to require a designer, a developer, a Figma file, and a week of back-and-forth now runs in your evening, by yourself, for free.</p>
<p>This is the part everyone has noticed. The part almost no one has noticed is what it means. The capability floor moved to the basement. The quality ceiling did not. The interesting question is not what got easier. It is what got harder.</p>
<h3>The loop is the same for everyone</h3>
<p>The mechanic is identical regardless of skill. Spec, generate, judge, refine, ship. A senior designer at Stripe runs this loop. A first-time founder running their personal site runs this loop. The model in the middle is the same. The IDE is the same. The output type is the same — real code that you can edit, deploy, and own.</p>
<p>What is different is what each person brings to the prompt and what each person notices about the output. The designer arrives with three reference sites pinned, a typographic system already in mind, and a vocabulary for what to push back on when the model defaults to safe-and-generic. The first-time founder arrives with &quot;make me a personal site&quot; and gets a personal site that is fine.</p>
<p>Both shipped. Only one shipped something memorable. The loop did not produce the gap. The inputs did.</p>
<h3>The four inputs that are not democratised</h3>
<p>Four things sit upstream of the prompt and decide how much of the loop&#39;s leverage you actually capture.</p>
<p>The first is taste. The skill of looking at three options and knowing which is better, and being able to articulate why. AI reverts to the mean. Reverting to the mean is what averages do. Calling it back to a sharper place requires someone who can see the better version and name the gap. Taste is years of looking at a lot of work, holding strong opinions about it, and being wrong publicly enough to recalibrate. There is no weekend version of this.</p>
<p>The second is reference fluency. Knowing what good looks like across enough categories to pull a relevant reference fast. Editorial layouts for memos. Product sites for SaaS. Portfolios for personal. Pasting three sites into Claude and saying &quot;this typography, this density, this restraint&quot; gives the model something to anchor on. Pure-text prompts default to bland. Most people skip the references because they do not have a library to pull from.</p>
<p>The third is spec discipline at minute one. The single biggest force-multiplier in the loop. A tight first prompt — type, palette, spacing scale, references, and the one thing the site is for — makes every subsequent prompt five times more effective. Most people treat the first prompt as throwaway and get throwaway results. The good designer-operator writes a real spec before they generate anything.</p>
<p>The fourth is just enough code vocabulary to be precise. Padding, line-height, max-width, gap. Not because you need to write it. Because being able to say &quot;tighten body line-height to 1.6, push max-width to 720px&quot; is faster than gesturing at vibes. A weekend gets you here. Most people do not spend the weekend.</p>
<blockquote>
<p>The AI compresses execution. It does not compress judgment. The judgment work moved to the front and to the back. Both require taste.</p>
</blockquote>
<h3>The exposed archetype</h3>
<p>The exposed archetype is the operator who saw the loop become free, concluded that design has been democratised, and started shipping. The output is fine. Everything is fine. Nothing is bad. Nothing is memorable either. The Webflow-template feel that used to come from a $19 template now comes from a free model, and the result is the same: indistinguishable from a thousand other sites built the same way, by people who brought the same nothing to the prompt.</p>
<p>This is not an argument that AI design tooling is bad. It is the argument that the tooling is now adequate, which means the variance in output is no longer about the tooling. The variance is about the operator. Five years ago, the constraint on shipping was that you could not get the loop to run. Now the constraint is whether you can run it well. Different problem. Same gap, just relocated.</p>
<h3>Personal evidence</h3>
<p>I shipped this site in a day with Claude. No design background. No Tailwind fluency before this week. No reference library worth the name. The loop ran. The site is fine.</p>
<p>A second version, after a weekend of staring at editorial sites and a weekend of CSS basics, would land in a different category. Not because I would have become a designer. Because the prompter would have gotten better. The model is the same model. The leverage is in the inputs, not the loop.</p>
<p>That is the experience of running the new loop on entry-level inputs. It works. It is also visibly not great. The gap between fine and great is the gap between the inputs I had and the inputs I did not.</p>
<h3>The implication</h3>
<p>The interesting question for the next decade is not whether AI replaces designers. It is what the new design hire is shaped like. Three years ago you hired Figma fluency, taste, and patience for handoffs. Now you hire taste, reference fluency, spec discipline, and basic code reading. The middle role — Figma jockey who hands off to a developer — is gone, the same way the mid-tier IC manager is gone. Both were translation layers in a system the model now translates inside of.</p>
<p>The democratised loop is real and worth using. It is also exposing a market that does not get talked about much yet — the gap between operators who run the same loop and produce wildly different outputs. The tools are not the moat. The taste is. The references are. The spec is. The model is doing the build. The work that used to be the build moved upstream and downstream, and both ends are still loaded with the things that take years to develop.</p>
<p>Anyone can ship a website now. Almost no one can ship a good one. The interesting question is which of those two operators you are about to become.</p>


AI democratised the design loop. It did not democratise the inputs.

Anyone can ship a website. Almost no one ships a good one.

<p class="lede">Most people setting up AI agents are underbuilding them and then concluding the agents do not work. They drop a one-liner into a chat, judge the first output, and walk away. The teams getting real leverage are doing something more thorough, and frankly more boring. They are treating agent setup the way they would treat onboarding a new hire. Heavy context, a clear manual, real tools, a feedback loop that actually closes.</p>
<p>The piece below is the manual that describes that process, in the order I would actually do it. It is a knowledge share, not an argument. Use it as a checklist when you are setting up your next agent in Cowork, Claude, or anything else.</p>
<h3>The six steps at a glance</h3>
<table>
<thead>
<tr>
<th>#</th>
<th>Step</th>
<th>What it produces</th>
</tr>
</thead>
<tbody><tr>
<td>1</td>
<td><strong>Map the workflow</strong></td>
<td>A doc that reads like an intern&#39;s manual</td>
</tr>
<tr>
<td>2</td>
<td><strong>Load the context</strong></td>
<td>An operator profile, a business profile, a calibration set</td>
</tr>
<tr>
<td>3</td>
<td><strong>Write guidelines and don&#39;ts</strong></td>
<td>Positive rules plus banned patterns</td>
</tr>
<tr>
<td>4</td>
<td><strong>Connect the tools</strong></td>
<td>Read access, write access, and clear guardrails</td>
</tr>
<tr>
<td>5</td>
<td><strong>Test and log misses</strong></td>
<td>A change log of what fell flat and why</td>
</tr>
<tr>
<td>6</td>
<td><strong>Update the system, not the prompt</strong></td>
<td>A versioned, living set of instructions</td>
</tr>
</tbody></table>
<p>The order matters. Each step assumes the previous one has been done properly. Skipping ahead is the most common reason setups feel mediocre and stay there.</p>
<h3>Step 1 — Map the workflow</h3>
<p>Before any prompting, write down the actual sequence of work the agent needs to do. Think of it as a manual you would hand to a smart intern on day one. What are the stages, where do decisions happen, where does the human stay in the loop, and what does &quot;done&quot; look like at each step. If you cannot describe the workflow clearly to another person, the agent will not infer it from a clever prompt, no matter how cleverly the prompt is written.</p>
<p>The first artefact in any good agent setup is a workflow document, not a prompt. The prompt comes later, and it gets shorter when the workflow is properly mapped, because most of the work has been done in the doc.</p>
<blockquote>
<p><strong>Checklist for Step 1</strong></p>
<ul>
<li><input disabled="" type="checkbox"> Stages of the work, in order</li>
<li><input disabled="" type="checkbox"> Inputs and outputs for each stage</li>
<li><input disabled="" type="checkbox"> Decision points where the agent should pause</li>
<li><input disabled="" type="checkbox"> Human checkpoints, with format for the handoff</li>
<li><input disabled="" type="checkbox"> Definition of &quot;done&quot; for the overall workflow</li>
</ul>
</blockquote>
<h3>Step 2 — Load the context</h3>
<p>The single biggest predictor of whether an agent&#39;s output feels right is the depth of context it has been given. Most setups starve the agent here, usually because the operator is being precious about token usage at the wrong stage. Setup is the one place where more is almost always better.</p>
<p>Three layers of context tend to matter the most. The operator profile, which covers who you are, how you think, what you care about, and how you talk. The business or domain context, which covers what you are working on, the recurring vocabulary, the metrics that matter, the constraints that shape your decisions. And a calibration set, which is a small library of best-in-class examples that show the agent the bar for &quot;good.&quot; The calibration set is the one most operators forget. It is also the one that makes the largest single difference.</p>
<pre><code>   CONTEXT LAYERS
   ─────────────────────────────────────────
   Operator profile    →  who you are, how you think
   Domain context      →  what you work on, vocabulary, constraints
   Calibration set     →  examples of best-in-class output
</code></pre>
<blockquote>
<p><strong>Checklist for Step 2</strong></p>
<ul>
<li><input disabled="" type="checkbox"> Operator profile (background, working style, communication preferences)</li>
<li><input disabled="" type="checkbox"> Domain or business context (vocabulary, metrics, constraints)</li>
<li><input disabled="" type="checkbox"> Calibration set (3-6 best-in-class examples of the desired output)</li>
<li><input disabled="" type="checkbox"> Explicit statement of the bar for &quot;good&quot;</li>
</ul>
</blockquote>
<h3>Step 3 — Write the guidelines and the don&#39;ts</h3>
<p>Positive instructions on their own are not enough. Models default to generic patterns when the instruction is broad, and the genericness is exactly what makes outputs feel mediocre. The work happens when you also name the failure modes you want to avoid.</p>
<p>In my experience the don&#39;ts do more work than the dos, especially for anything where voice or judgment matters. Banned phrases, structural patterns that read as slop, common mistakes from earlier outputs. The more specific the don&#39;ts, the more the agent&#39;s output starts to feel like yours rather than the model&#39;s defaults.</p>
<blockquote>
<p><strong>Checklist for Step 3</strong></p>
<ul>
<li><input disabled="" type="checkbox"> Positive rules (what good looks like)</li>
<li><input disabled="" type="checkbox"> Banned phrases and patterns (the AI tells you want to avoid)</li>
<li><input disabled="" type="checkbox"> Edge cases (when the standard approach does not apply)</li>
<li><input disabled="" type="checkbox"> Decision rule for when to ask the operator versus when to act</li>
</ul>
</blockquote>
<h3>Step 4 — Connect the tools</h3>
<p>An agent that can only talk is a thinking partner, which is useful but limited. An agent that can read your files, write to your storage, pull from your data sources, and push to your downstream tools is an operator. The leverage difference between the two is enormous, and most setups stop at the thinking-partner level because connectors feel fiddly to set up.</p>
<p>They are worth the time. A well-connected agent collapses entire workflows that used to require human-in-the-loop coordination. A poorly connected one stays a chat partner forever.</p>
<pre><code>   AGENT CAPABILITY LADDER
   ─────────────────────────────────────────
   Read only          →  thinking partner
   Read + write       →  drafting partner
   Read + write + push →  operator
</code></pre>
<blockquote>
<p><strong>Checklist for Step 4</strong></p>
<ul>
<li><input disabled="" type="checkbox"> Read access (what should the agent be able to see)</li>
<li><input disabled="" type="checkbox"> Write access (what is it allowed to modify or create)</li>
<li><input disabled="" type="checkbox"> Guardrails (what is off-limits, what requires explicit confirmation)</li>
<li><input disabled="" type="checkbox"> Confirmation points (where you want a human in the loop before action)</li>
</ul>
</blockquote>
<h3>Step 5 — Test and log the misses</h3>
<p>The first few outputs will be wrong in specific, recurring ways. The job is to spot the pattern in the misses rather than fix each one in the chat. Treating each miss as a one-off prompt fix is how setups stay mediocre, because the underlying instruction never improves and the same mistake will resurface in the next run.</p>
<p>Keep a running log of what fell flat and why. The log becomes the source of truth for system updates. Over a few cycles it tells you which file the fix actually belongs in, and which patterns are recurring rather than one-off.</p>
<blockquote>
<p><strong>Checklist for Step 5</strong></p>
<ul>
<li><input disabled="" type="checkbox"> What failed (specific output, specific run)</li>
<li><input disabled="" type="checkbox"> Why it failed (the underlying pattern, not just the symptom)</li>
<li><input disabled="" type="checkbox"> Which file the fix belongs in (workflow doc, context, guidelines)</li>
<li><input disabled="" type="checkbox"> What to test on the next run to confirm the fix landed</li>
</ul>
</blockquote>
<h3>Step 6 — Update the system, not the prompt</h3>
<p>This is the step most operators skip, and the one that decides whether the system compounds or stays flat. If the agent makes the same mistake twice, the fix belongs in the workflow doc, the context file, or the guidelines. It does not belong in a longer prompt the next time you talk to it.</p>
<p>Versioning the underlying system is what turns a useful agent into a sharp one over weeks and months. Treat the setup as a living artefact, not a one-off configuration. Bump the version when changes land. Keep a changelog at the bottom of every file. Never update silently, because silent updates make it impossible to debug regressions later.</p>
<blockquote>
<p><strong>Checklist for Step 6</strong></p>
<ul>
<li><input disabled="" type="checkbox"> Weekly or post-batch review of the change log</li>
<li><input disabled="" type="checkbox"> Proposed updates with the reason attached</li>
<li><input disabled="" type="checkbox"> Version bumped on the relevant file</li>
<li><input disabled="" type="checkbox"> Changelog entry written</li>
</ul>
</blockquote>
<h3>The compounding effect</h3>
<p>Each of these steps is unglamorous on its own. The payoff comes from doing all six and keeping the loop running. An agent that has a properly mapped workflow, deep context, named failure modes, real connectors, a miss log, and a versioned set of instructions gets sharper every week. An agent that has a clever prompt does not.</p>
<p>The setups that compound look more like internal documentation than like prompt engineering. That is the part most people miss. The manual you write before you need it is the one that keeps paying.</p>


A setup manual for AI agents that get sharper with use, not flatter

Six steps to an agent that compounds

<p class="lede">The line between craft and grind is fine. From the outside, both look the same. Both burn hours. Both look hard. Both produce something at the end of the day. The felt difference is invisible to everyone, including you, unless you stop and look at it.</p>
<p>This is a note on stopping to look.</p>
<h3>Why this matters</h3>
<p>The metrics that make up modern working life are louder than they have ever been. Hours posted. Tickets closed. Output shipped. Lines added. Every dimension of the work that is countable has been instrumented, ranked, and made visible. The dimensions that are not countable, direction, refinement, the specific thing you are getting better at, are quieter than they have ever been.</p>
<p>The gap between what is loud and what matters is widening. Most people are taking the loud thing for the real thing. It is the easier mistake to make and the harder one to notice.</p>
<p>This is the part that is uncomfortable. The cost of grind is not the hours. The hours come and go. The cost is direction. A year spent grinding in the wrong direction is a year that cannot be respent. Opportunity cost is the only cost that matters at this timescale, and it is the one you cannot see while it is happening.</p>
<div class="pull-quote">Opportunity cost is the only cost that matters at this timescale, and it is the one you cannot see while it is happening.</div>

<p>Speed is celebrated regardless of direction. The most expensive thing in a career is to be very fast and very lost. You arrive at a place you did not choose, having spent years getting there. By the time you notice, the years are gone and the direction is locked.</p>
<p>Craft is the opposite shape. Slower. Quieter. Less performable. The satisfaction it produces is a different kind of satisfaction. It does not feel like winning the day. It feels like a thing slowly becoming the thing it should be. Most people running hard right now do not have access to that feeling. They have replaced it with the calendar.</p>
<h3>The questions worth sitting with</h3>
<p>So the question worth sitting with, quietly, with no one watching, is not whether you are working hard. It is whether you are walking somewhere.</p>
<p>Five questions follow. They are uncomfortable on purpose.</p>
<p><strong>What are you walking towards?</strong> Not &quot;what is your goal.&quot; That is the answer everyone has rehearsed. The specific thing your years are pointed at. The thing that, if you stopped today, would still be visible in the work.</p>
<p><strong>Could you describe what you have been getting better at, in a sentence, without using the word &quot;more&quot;?</strong> &quot;More productive&quot; does not count. &quot;More efficient&quot; does not count. The specific craft, the specific bar, the specific thing that is sharper now than it was a year ago. If the sentence does not come, the sentence does not exist.</p>
<p><strong>When you imagine yourself five years from now, are you doing more of what you do now, or different work that has required you to grow?</strong> The honest answer is uncomfortable for most people running hard right now. The grind tends to produce more of the same person, faster.</p>
<p><strong>What did you say no to this week to protect the direction you care about?</strong> If the answer is nothing, the direction is not protected, which means it is not real. A direction that is never defended is a preference you have about yourself, not a place you are walking.</p>
<p><strong>If all the metrics vanished tomorrow, the hours, the ranks, the titles, the salary, what would still be true about you that you would be proud of?</strong> That residue is what craft produces. Grind does not produce a residue. It produces a calendar.</p>
<div class="pull-quote">Grind does not produce a residue. It produces a calendar.</div>

<h3>The walk and the hours</h3>
<p>The questions are uncomfortable on purpose. Hours are easy to count. Direction is not. The gap between the two is the gap between five years of grind and five years of craft.</p>
<p>The hours are not the work. The walk is.</p>


A note on the fine line between craft and grind

You can be very fast and very lost

<p class="lede">The most honest reading of grindset is not that it produces worse work. It is a signal, and the signal works because nobody can see anything else.</p>
<p>Most knowledge work is illegible. Nobody outside your team can tell whether the deck was good. Nobody outside the cap table can tell whether the strategy was sound. Judgment is private. Effort is public. That asymmetry is not incidental to grindset. It is the whole reason grindset exists.</p>
<h3>Effort is the cheapest credential</h3>
<p>In any market where the output is hard to evaluate, actors substitute easier metrics. Hours. Outputs shipped. Visible discomfort. The 4am alarm is not productivity.</p>
<p>Signalling.</p>
<p>The cheapest credential available to anyone who cannot show the more expensive one, which is taste.</p>
<p>The reason it works as a signal is that the cost is real. Sleep deprivation. Missed weekends. The bloodwork panel that says you are running too hot. Costly signals carry information, because nobody fakes a cost they do not have to pay. That is the part the grindset class understands implicitly and never says out loud. The cost is the credibility.</p>
<div class="pull-quote">Effort is public. Judgment is private. That is the whole game.</div>

<p>Grindset is loudest in industries where the bar for taste is high and the legibility of taste is low. Founders. Consultants. Investors. Bankers. The places where everyone knows judgment matters and nobody has a clean way to measure it from the outside. So the credential becomes hours. The credential becomes intensity. The credential becomes the screenshot of the inbox at zero, the calendar full of meetings, the LinkedIn post about the morning routine. The credential is anything that costs you something the next person is not paying.</p>
<h3>What the signal actually costs</h3>
<p>The catch is who pays. The signal is socially valuable. The cost is privately catastrophic.</p>
<p>The grindset class is more anxious, more sleep-deprived, and more medicated than the generation before it. Cortisol elevated. SSRIs prescribed. GLP-1s repurposed as work accessories. The Whoop scores trending in the wrong direction year on year. The sleep tracker, the very tool of the optimisation cult, telling you the optimisation is failing, and the answer always being more optimisation. The body keeps the receipts. The signal does not.</p>
<p>This is not the same dynamic as a CEO pushing a winning team, or a coach asking more of an elite athlete. That is a different game, with a different accountability structure, and the person doing the pushing has line of sight into whether the push is working. Grindset is not someone being coached against a clear standard. It is someone running themselves into the ground because effort is the only signal they have left to send and they cannot afford to stop sending it.</p>
<p>The exposed archetype is the operator without taste or acumen using volume as a credential. Visible everywhere. Loudest in their thirties. Increasingly hard to distinguish from a person actually doing the work, until the work has to ship and the difference is finally legible.</p>
<h3>The signal is breaking</h3>
<p>The arbitrage that built grindset is closing. The reason is unglamorous. Judgment is becoming legible faster than grindset&#39;s defenders can adapt to it.</p>
<p>AI is the proximate cause. When two people start with the same brief, the same data, and the same access to a model, the difference between their outputs is not hours. It is taste. And taste is becoming visible precisely because the inputs are being held constant. The deck either works or it does not. The strategy either compounds or it does not. The hiring decision either ages well or it does not, and the timeline on &quot;ages well&quot; is shrinking from years to months as feedback loops compress.</p>
<p>This is the part the 4am alarm class does not see coming. They built a career on a signal that was valuable because judgment was private. The signal stops working when judgment goes public. The cold plunge will keep working as a personal habit. It will stop working as a career credential. The same is true of the inbox at zero, the meeting count, the bragged-about hours.</p>
<div class="pull-quote">The grindset trade was always temporary. It just took longer than expected for the trade to close.</div>

<p>Grindset was never a productivity ideology. It was an information-asymmetry trade dressed up in masculine language about discipline. The trade is closing on its own. The people who built their careers on it are about to discover what their work actually looked like underneath, once the effort signal stops carrying the weight it used to. That bill is being paid in cortisol and credentialism. The second one is harder to recover from than the first.</p>


What it actually buys, what it costs, and why the trade is about to close

The grindset is a signal, not a strategy

<p class="lede">The story sells itself in two flavours. The first says you should grind for forty years, defer everything that matters, and trust that the life will still be there waiting when you finally reach it. The second says you should blow it all up at twenty-five, take the nine-month trip, &quot;find yourself,&quot; and figure out the rest later. Both versions get told as if they were the only two on offer. They are not, and the trick the story is playing is to make the wide space between them feel invisible.</p>
<p>The reason the dichotomy keeps working, given how badly it fits the lives most people actually live, is that the romantic version of total commitment is older than any of us. The struggling-artist trope did not really die; it just changed clothes. The painter who starves for the work has been replaced by the founder who proudly mortgaged the house, the creator who quit the corporate job with no audience, the traveller who turned the gap year into a personal brand. The costumes are different and the implicit claim is identical. Anything less than total commitment is a failure of nerve.</p>
<p>The reason that claim keeps thriving is not that it describes reality particularly well. It is that it makes good content. A clean before-and-after, the cubicle to the camper van, the spreadsheet to the studio, fits a ninety-second video and an algorithm that rewards extremes. The middle, where most people actually live, does not.</p>
<h3>What the middle actually looks like</h3>
<p>Almost every life worth living, watched at close range, gets built somewhere in that wide grey middle. There is the person who keeps the day job and writes seriously on the side for years before anything publishes. The operator across the corridor who saves carefully and builds skills nobody is paying them to build yet, and takes the real risk only when the time arrives, looks like a hedge until they stop being one. The artist with a partner and a mortgage and a body of work that has compounded across a decade has been able to compound that work precisely because the rest of their life was stable enough to hold it. None of these people show up on your feed. The story is too quiet, too long, too unsegmentable, and quietness is what the algorithm cuts first.</p>
<p>The two visible archetypes in the dichotomy, the twenty-five-year-old who blew their savings to find themselves and the thirty-five-year-old still in a soul-flat job because they over-hedged, are visible because they are extreme, not because they are common. Confusing visibility for frequency is how the false dichotomy keeps its grip. The middle is the actual baseline. We just do not see it, because the middle is not the kind of life that gets photographed.</p>
<h3>Real passion, watched closely, is quieter than the trope</h3>
<p>What anyone who has watched a passion last a decade tends to notice is how unlike the trope it looks. The pattern is long-term focus on something specific, often unfashionable, with no obvious payoff for years at a time. There is reflection on what is and is not working, sometimes in private journals, sometimes in conversations with a small number of people who care about the same questions. There is patience with the parts of the work that take years to compound, and a willingness to keep going through the seasons when nothing is photogenic and nobody is clapping. Recklessness, which the trope sells as the substance of passion, turns out to be its cosplay. The substance is much quieter, and much harder to keep up.</p>
<div class="pull-quote">Recklessness is the cosplay of passion, not its substance.</div>

<p>The same surprise runs in the other direction when you look closely at what responsibility actually means in those same lives. It is not the cardboard version, where the responsible person is grey and dead inside and quietly resentful of their own choices. It is something more interesting, and less photogenic, than that. They kept their skills sharp through stretches when the work was not glamorous. They built savings that bought them runway when an opportunity finally arrived. They looked after their bodies and their relationships, because the slow work of caring deeply about a craft consumes a person, and a thin life cannot hold it. Most people who burn out on a creative project do not burn out because the work was too hard. They burn out because the rest of the life was too thin to hold the work.</p>
<p>The dichotomy collapses on inspection. What looks from a distance like the reckless artist usually has a quiet structure underneath, often someone else&#39;s structure they have borrowed without naming. The version of the responsible person who has made peace with being grey and dead inside almost always has private, persistent obsessions running in the background, kept small only because nothing in the life is set up to support them growing. The visible part of the costume is rarely the part doing the load-bearing work.</p>
<h3>What this means for the choice you actually have</h3>
<p>The choice in front of most people, once the trope has been stripped away, is not between grinding for forty years and blowing it all up at twenty-five. There is a quieter and more demanding question underneath. How do you build the scaffolding now, while you have the time and the energy to build it, so that the things you actually care about have somewhere to live for the next thirty years.</p>
<p>The honest answer to that is unglamorous, and it is the kind of answer that almost nobody is going to share a video about. It involves staying employable in work you can stand, keeping enough optionality that you can still change direction when something interesting becomes possible, building savings without making them the point of the exercise, taking care of a body you have to live inside, and investing seriously in the small number of relationships that actually carry weight when life leans on them. None of that fits a clean before-and-after frame. All of it is the kind of thing that, twenty years in, turns out to have been the work all along.</p>
<p>Pick both. The boring half is what makes the interesting half last.</p>


The struggling-artist trope has shape-shifted, and the dichotomy is still a lie

Responsibility and passion are not opposites

<p class="lede">The most expensive virtue I was taught is humility.</p>
<p>It was the internal kind, the one that tells you to wait until you are ready, post when your thinking is bulletproof, move overseas when your language is fluent, email the leader once you have something to say. Calibrate yourself accurately. Earn your way in. Do not get ahead of your evidence.</p>
<p>That is good advice for staying respectable. It is bad advice for getting where your abilities should have put you a long time ago.</p>
<p>I am writing this from inside the trap. I have spent most of my twenties playing a game I am quietly very good at, in rooms I am quietly the smartest person in, watching people I am pretty sure I am no worse than do bigger things, in better cities, with bigger jobs, earlier. The thing I want to argue is that this is not unfair. It is the price I have been paying for a long time without knowing what the bill was for.</p>
<p>The bill is the audacity tax.</p>
<h2>The choices that look reasonable in hindsight</h2>
<p>I picked UCLA over Tokyo because I was nervous about the language. I picked BCG and a hopeful internal transfer over moving cold to Asia because the transfer felt like a clean line on a CV. I picked Southeast Asia for my travel year over Europe or South America because I knew people, the food was familiar, and the flights were short. I picked staying in my friend group over the events I felt I did not deserve to be at. I picked never posting publicly over posting half-formed thinking. I picked never emailing the leaders I admired over being briefly, embarrassingly, ignored.</p>
<p>Every single one of those choices was defensible to anyone who asked. That is the part that took me too long to see. The pattern is not risk aversion. The pattern is choosing, every time, the option I could justify on paper over the option that would have widened my range.</p>
<p>The legible choice has a hidden cost. The cost is not the wins I missed. The cost is the version of me that would have been built by being out there.</p>
<h2>The shape of the tax</h2>
<div class="pull-quote">The audacity tax is what bright, calibrated people pay for choosing the defensible option over the radius-extending one.</div>

<p>It is invisible because the bill never arrives. You do not get a notification when you miss a job, a friendship, a city. You just quietly are not where your abilities suggest you should be. The numbers do not move, the story stays clean, the CV looks fine. Five years pass. Ten years pass.</p>
<p>The cost compounds in three places, and I think this is the part most of the existing writing on this misses. It is not just the wins.</p>
<div class="framework-grid">
  <div class="framework-card">
    <div class="framework-letter">1</div>
    <div class="framework-name">Failures you never had</div>
    <div class="framework-desc">Which means lessons you never internalised.</div>
  </div>
  <div class="framework-card">
    <div class="framework-letter">2</div>
    <div class="framework-name">People you never met</div>
    <div class="framework-desc">Because you were not in the rooms where they were.</div>
  </div>
  <div class="framework-card">
    <div class="framework-letter">3</div>
    <div class="framework-name">Mentors who never picked you</div>
    <div class="framework-desc">Because you were never visible enough to be picked.</div>
  </div>
</div>

<p>That is what extends radius. Failures on the record. People you embarrassed yourself in front of. Versions of you that did not work and that you got to bury and try again on.</p>
<p>Calibration does not compound. Audacity does.</p>
<h2>I was waiting for the timeskip</h2>
<p>I think a lot now about what I was doing all those years.</p>
<p>If your reference points are stories, you know the trick. The protagonist hits a wall. The protagonist goes off to train. Two years pass in a montage, or a hard cut, or a screen that just says &quot;two years later.&quot; The protagonist comes back ready. The world has been waiting. The skills are stacked. The next fight is winnable.</p>
<p>I have lived a version of this story in my head for a decade. Get the right grades, get the right job, read the right books, build the right CV. Calibrate. Train. Then, one day, emerge. The world will be roughly where I left it and I will be ready for it.</p>
<p>Real life does not give you the timeskip. There is no two-year cut where the protagonist is the only one moving and everyone else holds still. Every year I was training in private was a year other people were out there, trying things, missing, getting picked, getting good in public. Skills built in isolation only compound when they meet the world. Mine were not meeting the world.</p>
<p>The version of this that is most painful to sit with is that the people I have envied, the ones who got jobs I did not think they deserved and the cities I would have liked, are not better than me. They are not smarter than me. They were just willing to be visibly wrong, in public, while I have been waiting for my arc to start.</p>
<div class="pull-quote">The audacious are not braver in a way I cannot reach. They are playing a higher-volume game I opted out of.</div>

<h2>Three questions before you keep paying</h2>
<p>The reason I have not landed this argument as theory is that the people I want it to reach are smart enough to argue with theory and not be moved by it. So three questions, and the cost only registers when you sit with your own answers.</p>
<p><em>When have you felt envy at someone you believe you are better than, and what did they actually do that you did not?</em></p>
<p><em>When have you missed a shot you knew you could have taken, and what did that quiet &quot;no&quot; actually cost you?</em></p>
<p><em>How long are you willing to keep paying this tax?</em></p>


On why bright, calibrated people quietly underwrite their own ceiling

The audacity tax

<p class="lede">The AI creative discourse has settled into two camps. AI will replace creative teams. Or AI slop will collapse engagement and we should retreat to human craft. Both are wrong because they are arguing about the wrong axis.</p>
<p>The actual question is where on the speed-vs-taste curve a given piece of creative should sit. The answer changes by funnel stage, by brand positioning, and by business size. There is no single right answer for a marketing team. There is only a right answer for a specific decision.</p>
<h2>Two metrics are diverging</h2>
<p>Engagement rates on AI-generated creative are softening. Hook rates, video view-through, CTR. That is real, and it is accelerating. But conversion rates on templated, tested formats are holding, and in many cases improving, because AI lets one person produce ten variants where they used to produce one. The funnel further down is fine. The teams panicking about AI creative are looking at one metric and missing it.</p>
<div class="ascii-chart"><pre>     Indexed performance, last 12 months

<p>  ▲
  │  ╲
  │   ╲╱╲                                          Engagement
  │      ╲╲                                        (CTR, hook rate)
  │        ╲╲╱╲
  │            ╲╲╲___
  │                  ╲╲___
  │
  │       _________________________  Conversion
  │  ___╱╱                            (CVR, ROAS, CPA)
  │
  └────────────────────────────────────────────────▶
     M1   M3   M5   M7   M9   M11   today</pre></div></p>
<p class="ascii-caption">[chart: two-line divergence. Engagement metrics trending down. Conversion metrics flat or up. The AI discourse is reading the wrong line.]</p>

<h2>Where AI creative wins</h2>
<p>Anything with a repeatable, replicable format. The structure carries most of the persuasion. Variation just needs to be relevant.</p>
<ul>
<li>Discount mechanics. Urgency. Comparison ads. Before-after. Problem-solution. Proof-point ads with rotating data.</li>
<li>Lo-fi statics. Scripted UGC reads. Anything you could brief in a sentence and produce a hundred versions of without losing the point.</li>
</ul>
<p>A Mounjaro 5mg lo-fi run earlier this year collapsed our CAC in a week. The format was tested. The variation was AI-fast. It worked because we were not asking AI to be original. We were asking it to be relentless inside a known frame.</p>
<h2>Where AI creative cannot go</h2>
<p>The work that requires a point of view. Breakthrough hooks. Virality. Mid-funnel emotional connection. Trust at the brand level. The visceral reaction that makes someone remember a brand a week later.</p>
<p>These outcomes need taste, and taste is exactly what AI reverts to the mean on. The more premium the brand, the more this matters. The more saturated the category, the more this matters. The more your edge depends on people feeling something specific about you, the more it matters.</p>
<h2>The framework: speed-taste matrix</h2>
<p>Map two axes. Funnel stage on one. Brand positioning on the other. The right mix of AI and human craft is different in every cell.</p>
<div class="ascii-chart"><pre>                    AI suitability matrix

<pre><code>                PREMIUM     MAINSTREAM     VALUE
              ┌─────────────────────────────────────┐
   TOP        │  ░░░░░░░  │  ░░░░░░░░  │  ▒▒▒▒▒▒▒  │
   FUNNEL    │  taste-led │  taste-led │  mixed    │
              ├─────────────────────────────────────┤
   MID        │  ░░░░░░░░ │  ▒▒▒▒▒▒▒▒  │  ▓▓▓▓▓▓▓  │
   FUNNEL    │  taste-led │  mixed     │  AI-led   │
              ├─────────────────────────────────────┤
   BOTTOM     │  ▒▒▒▒▒▒▒  │  ▓▓▓▓▓▓▓▓  │  ████████ │
   FUNNEL    │  mixed     │  AI-led    │  AI-heavy │
              └─────────────────────────────────────┘
          ░ low AI suitability    █ high AI suitability&lt;/pre&gt;&lt;/div&gt;
</code></pre>
<p class="ascii-caption">[matrix: 3x3 grid. Bottom-funnel × value = highest AI suitability. Top-funnel × premium = lowest. The middle is where most teams sit and where most teams lose.]</p>

<h2>The metrics that should sit under each cell</h2>
<ul>
<li><strong>Bottom-funnel templated work.</strong> CPA, ROAS, CVR. Volume and iteration count matter. AI wins.</li>
<li><strong>Mid-funnel work.</strong> Brand recall lift, engaged view-through, assisted conversion. Mixed. AI helps with variation. Humans set direction.</li>
<li><strong>Top-funnel breakthrough.</strong> Organic share rate, branded search lift, qualitative resonance. Human-led.</li>
</ul>
<p>The teams winning right now run a heavy-on-both-ends shape. Volume of AI-templated work at the bottom of the funnel. A small number of human-led hero pieces at the top to anchor the brand. The thing that hollows out is the middle. Mid-quality, mid-effort, mid-distinctive output. Generic enough that AI could do it. Slow enough that AI is doing it faster. Neither speed wins nor taste wins.</p>
<div class="pull-quote">The hard part is not picking the mix. It is keeping the learning loop fast enough to compound while still holding a point of view distinct enough to not regress to the AI mean. Taste at the prompt level, not the asset level.</div>

<h2>The question worth asking</h2>
<p>The question worth asking your team is not whether AI replaces creatives. It is which creative decisions you are willing to template, and which ones you would defend with your career. The second answer is the part of your team that needs to get sharper, not smaller.</p>


Where AI creative actually wins, and where it cannot

The speed-taste tradeoff

<p class="lede">Most marketing teams are measuring AI transformation with the wrong metrics. They count assets shipped. Hours saved. Cost per output. Both framings miss the point entirely.</p>
<p>Output and efficiency are not outcomes. They are symptoms of something deeper, or methods that produce something else. Treating them as goals is procurement dressed as strategy.</p>
<p>The actual outcome is speed to learning. The consequence is faster, higher-quality business decisions. That is the thing AI changes when it is working. Everything else is a side effect.</p>
<h2>What I learned scaling a DTC brand</h2>
<p>Running performance and creative at Juniper Japan over the last year reframed how I think about this. The wins that mattered were not the ones where AI made the team produce more. They were the ones where AI compressed the time between forming a hypothesis and knowing whether it was right.</p>
<p>A specific week earlier this year is the cleanest example. Multiple things landed in seven days. A price drop on Mounjaro 5mg. A new squeeze page on Google. A lo-fi static creative refresh on Meta. A fold reduction on the landing page. Each of those individually was a hypothesis. CAC fell from ¥107K to ¥43K in a week.</p>
<p>The reason they compounded into that result was not that AI made any one of them faster to produce. It was that the entire feedback loop, from diagnose to hypothesise to build to ship to read to decide, collapsed from weeks to days. We were not producing more assets. We were producing more learning.</p>
<div class="pull-quote">When I look at what changed, it was not the tooling. It was the length of the loop.</div>

<h2>Why output and efficiency are the wrong outcomes</h2>
<p>Output and efficiency are easy to measure, which is why they get adopted as KPIs. Measuring them as outcomes creates two failure modes.</p>
<p><strong>Output as outcome</strong> rewards volume regardless of whether the volume is generating learning. Teams produce more creative, more emails, more landing pages, and learn nothing faster than before, because the loop was never the bottleneck. Decision velocity was.</p>
<p><strong>Efficiency as outcome</strong> rewards cost reduction, which in a growth function is almost always the wrong objective. A marketing team running at fifty percent lower cost but the same learning rate has not transformed. It got cheaper.</p>
<p>Both framings treat AI as a tool for doing the same thing more cheaply. That is not transformation. That is procurement.</p>
<h2>The metrics that actually matter</h2>
<p>A team that has properly absorbed AI shows it in four places, none of them about output:</p>
<ul>
<li><strong>Hypothesis cycle time.</strong> Idea to validated learning. Should compress 3-10x.</li>
<li><strong>Decision throughput.</strong> Meaningful business decisions per week, without quality loss.</li>
<li><strong>Kill speed.</strong> How fast bad bets get shut down.</li>
<li><strong>Compounding institutional knowledge.</strong> Whether each loop leaves a trail the next loop can stand on.</li>
</ul>
<p>None of these are output metrics. None are efficiency metrics. They are properties of the learning system itself.</p>
<h2>The ambition gap</h2>
<p>The reason most AI transformation programmes underwhelm is that they are scoped against the wrong outcome. &quot;We will save forty hours a week&quot; is a procurement pitch dressed up as transformation. &quot;We will cut our hypothesis cycle from three weeks to three days and double the number of validated business bets we make per quarter&quot; is transformation.</p>
<div class="pull-quote">If your AI strategy reads like a list of efficiencies, you have built a cost-cutting plan. If it reads like a redesign of how decisions get made and how fast learning compounds, you have built a competitive advantage. Most plans are the first thing wearing the second thing's clothes.</div>

<p>Pick the second one. The first one is what every other team is doing, and being marginally cheaper than them is not a strategy.</p>


On why output and efficiency are the wrong outcomes for AI in marketing

Speed to learning is the only metric that matters

<p class="lede">The dominant fear about AI in knowledge work is that it eats the bottom of the org chart. Juniors out, entry-level roles hollowed, the grunt work absorbed first. It is a tidy story, and it is looking at the wrong layer. The real squeeze is happening higher up, and most middle managers have not noticed yet.</p>
<p>The middle layer is getting compressed, but not because managers do not matter. The part of the manager job that AI most easily absorbs is exactly the part that bloated the middle in the first place. Status reporting. Project coordination. Information relay. The work that exists because information moves slowly through a hierarchy. Strip that out and you do not lose management. You lose a layer of management that was mostly translation.</p>
<p>The compression is structural. AI does not just shrink roles and skill sets. It shrinks the hierarchy itself. Decisions that used to need three layers of context-passing can now happen in a single conversation. That is the change worth naming, and it is doing more work than most of the discussion gives it credit for.</p>
<h2>A quick look back</h2>
<p>Some history is worth holding in mind. The 1980s and 90s saw waves of delayering, driven by cost. The 2000s and 2010s rebuilt the middle as companies scaled and complexity returned. What is happening now is the third turn of that cycle, but the forcing function is different. It is not cost this time. It is speed. Decisions need to move faster than a hierarchy can carry them, and the layers that exist primarily to move information are the first to look expensive.</p>
<h2>What replaces the pure manager</h2>
<p>What replaces the pure manager is the IC manager. Someone who can swap between managing and doing without it feeling like a regression. Someone who refines their craft alongside their team rather than above it. Someone who stays close enough to the details that their strategic calls are grounded in real texture, not in summaries that have been polished three times before reaching them.</p>
<p>The IC manager does not have to be better than their team at any specific craft. They have to be good enough to make high-quality judgment calls on the fine details. Good enough that when they push back on a piece of work, the pushback lands. Good enough that the strategy they shape is informed by the actual shape of the customer experience, the internal workflow, the small pain points that never make it into a status update because they are too granular to bother reporting.</p>
<h2>Who is exposed</h2>
<p>The exposed archetype is the pure strategic thinker. The manager who talks well, frames things well, says a lot of things in meetings, and never touches the work. Corporate inertia protected this role for a long time, partly because the information advantage that came from sitting in many meetings and synthesising across them was real. That advantage is now available to anyone with a transcript and a model. The pure strategic thinker who never touches the craft is going to find their unique value harder to point to than it used to be.</p>
<div class="pull-quote">The information advantage that came from sitting in meetings and synthesising is now available to anyone with a transcript and a model.</div>

<h2>The reframe on &quot;lower leverage&quot; work</h2>
<p>The conventional view is that managers should delegate execution and focus on leverage. The IC manager view is that staying close to the texture is the source of the leverage. The work that gets called &quot;lower leverage&quot; — the hands-on building, the actual customer call, the specific piece of copy — is in fact the input to the high-leverage decisions. Strip it out and the decisions get worse, not because the manager is unintelligent, but because they are working off a flattened picture.</p>
<p>The best managers in the next three years will be the ones who can drop into the work, ship something themselves, and come back up to the strategic altitude with a sharper view. Not because doing the work is romantic. Because the calibration loop is what makes the judgment trustworthy.</p>
<h2>The implication for how to design teams</h2>
<p>Fewer layers. More IC managers. A lot less tolerance for roles that exist primarily to translate between layers. The manager who survives the next three years is the one who is also building. The one who is only directing is being outpaced by the one who is doing both, and the gap is going to widen rather than narrow.</p>
<div class="pull-quote">The manager who survives the next three years is the one who is also building.</div>

<p>If you are designing a team right now and your org chart has a middle layer that exists to relay information between strategy and execution, that layer is borrowed time. If you are managing right now and you have not touched the work in a year, that is a flag worth taking seriously. Not as a sign of failure, but as a sign that the role you are in was designed for a different speed of decision-making than the one that is coming.</p>


Why AI is squeezing the middle of the org chart, not the bottom

The rise of the IC manager

<p class="lede">Every job description in 2026 says &quot;AI-native.&quot; Almost none of them define it. The hiring market is full of vague language, vibes-based interviews, and rubrics that test for tool fluency instead of judgment. AI-native, as it stands today, is theatre. People know they should want it. They cannot describe what it is.</p>
<p>The lazy version of the answer is &quot;they use ChatGPT well.&quot; That is a symptom, not a trait. Tools change every six months. The traits that actually predict whether someone thrives in a post-AI environment are stable across whatever the next tool is. They sit underneath the tools. They are the things you should be looking for in a candidate, in a trial, in a reference check, in your own people.</p>
<p>There are six. They spell POST-AI. The acronym is a memory aid. The six traits are the rubric that should sit underneath every &quot;AI-native&quot; job description in the world right now, and almost never does.</p>
<h2>The framework, at a glance</h2>
<div class="framework-grid">
  <div class="framework-card">
    <div class="framework-letter">P</div>
    <div class="framework-name">Prioritisation</div>
    <div class="framework-desc">Choosing what matters when you can do anything.</div>
  </div>
  <div class="framework-card">
    <div class="framework-letter">O</div>
    <div class="framework-name">Ownership</div>
    <div class="framework-desc">Holding outcomes through complexity AI cannot yet carry.</div>
  </div>
  <div class="framework-card">
    <div class="framework-letter">S</div>
    <div class="framework-name">Systems</div>
    <div class="framework-desc">Designing workflows that compound, not just doing tasks.</div>
  </div>
  <div class="framework-card">
    <div class="framework-letter">T</div>
    <div class="framework-name">Taste</div>
    <div class="framework-desc">Knowing what is good when generation is free.</div>
  </div>
  <div class="framework-card">
    <div class="framework-letter">A</div>
    <div class="framework-name">Alignment</div>
    <div class="framework-desc">Trust, relationships, and the ability to move information cleanly.</div>
  </div>
  <div class="framework-card">
    <div class="framework-letter">I</div>
    <div class="framework-name">Insight</div>
    <div class="framework-desc">The sharpness of the inputs that go into the decision.</div>
  </div>
</div>

<p>The deep version of each, below.</p>
<h2>P — Prioritisation</h2>
<p>When the cost of doing anything has collapsed, the cost of doing the wrong thing has not. The constraint shifts from execution to selection. The trait worth hiring for is the ability to ruthlessly choose what matters out of what is possible, and to be willing to drop a lot of things that look attractive to get there.</p>
<p>This is harder than it sounds. Most people are good at adding to a roadmap. Very few are good at cutting one in half and defending the cut. The candidate who can do that is the candidate you want.</p>
<h2>O — Ownership</h2>
<p>AI can execute almost any task. It cannot yet own an outcome through complexity, cross-functional handoffs, edge cases, and downstream consequences. The trait worth hiring for is the willingness to take a thing end to end. Fix it when it breaks. Notice the small things upstream and downstream. Hold the result, including the parts that fall outside whatever the original brief said.</p>
<p>The exposed candidate here is the one who shipped their portion of the project and pointed at someone else for everything that came after. AI makes that person less useful, not more, because the seams between functions are where the real failures happen.</p>
<h2>S — Systems</h2>
<p>The shift from doing tasks to designing systems that do tasks. Workflows that self-reinforce. Learning loops that compound. The ability to look at a function, redesign it as a system that orchestrates agents and humans, and then supervise the system rather than the work. This was always a senior trait. AI makes it a baseline trait, even at junior levels, because the alternative is being the person AI replaces task by task.</p>
<h2>T — Taste</h2>
<p>When generation is free, the bottleneck moves to judgment of what is good. What is distinct. What is worth keeping. Taste is the most under-discussed of the six because it is the hardest to test for, but it is the trait that prevents an organisation from drowning in its own output.</p>
<p>You cannot teach taste in an interview. You can test for it. Show three pieces of work. Ask the candidate which is best and why. The reasoning matters more than the choice.</p>
<h2>A — Alignment</h2>
<p>Trust is the spine of this one. The ability to build durable, trusted relationships that move information faster than any tool, sharpen decisions through honest pushback, and pull people toward what actually matters. AI raises the noise floor on every channel. Alignment is what cuts through it. The person who has earned trust across functions makes better decisions and gets better information than the person who has not, regardless of how good their tools are.</p>
<div class="pull-quote">AI raises the noise floor on every channel. Alignment is what cuts through it.</div>

<p>This is the trait recruiters under-test for, because it does not show up on a CV and it is awkward to ask about directly. The proxy is reference checks done seriously. Not &quot;would you hire them again.&quot; Specific questions, with people who worked cross-functionally with the candidate, about how they handled disagreement, surfaced bad news, and earned the trust of people who had no reason to give it.</p>
<h2>I — Insight</h2>
<p>The quality of the inputs you bring to the decision. Customer feedback. Patient research. The texture of the actual problem. AI does not improve the quality of inputs. It compounds whatever you put in. The person with sharper insight makes the AI-augmented decision better. The person without it makes faster bad decisions.</p>
<p>Insight is what separates the operator who runs ten experiments and learns nothing from the operator who runs three and changes the company. Both moved fast. Only one moved correctly.</p>
<h2>The thing that is missing, on purpose</h2>
<p>Speed.</p>
<p>Everyone wants to hire for speed. Speed is not a trait. It is an outcome. The fast operators in 2026 are fast because they prioritised correctly, owned the work end to end, built systems, applied taste, were aligned with the right people, and had real insight into the problem. Strip any of those out and the speed is just slop, faster.</p>
<p>Hiring rubrics that optimise on speed and efficiency directly are the fastest way to fill a team with slop cannons. People who ship a lot of mediocre work. People who confuse motion with progress. The trial that asks &quot;how many things can you do in two hours&quot; selects for exactly the wrong thing. The trial that asks &quot;what would you cut and why&quot; selects for the right one.</p>
<div class="pull-quote">Speed is not a trait. It is an outcome.</div>

<h2>Different roles weight the six differently</h2>
<p>No one is hireable on all six. The role decides the priority. A few worth naming.</p>
<p><strong>Designers and creatives.</strong> Taste leads. Insight close behind. Without taste the role is one AI can already approximate. With taste, plus the ability to direct AI tooling, the role is more leveraged than it has ever been.</p>
<p><strong>Product managers.</strong> Prioritisation and Ownership lead, by a long way. PMs become more valuable, not less, as engineering speeds up. The cost of building the wrong thing accelerates with execution speed, which means the cost of bad prioritisation accelerates too. The PM who cannot say no is now actively dangerous, in a way they were not when shipping was the bottleneck.</p>
<p><strong>Engineers.</strong> Systems leads. Ownership close. Taste matters more than people pretend, especially at the senior end. The engineer who can architect for AI-augmented teams, hold the outcome through deployment, and make taste-led calls on what to build is the engineer compounding fastest.</p>
<p><strong>Marketers and growth operators.</strong> Insight and Taste lead. Systems for anyone running paid media or lifecycle. The marketer who has texture on the customer and the discipline to act on it, paired with the systems mindset to scale what works, is irreplaceable. The marketer who only knows tools is a slop cannon waiting to happen.</p>
<h2>What this means for how to hire</h2>
<p>Stop asking &quot;do you use AI&quot; in interviews. The question is theatre and the answer is meaningless. Build trials that test for these six instead.</p>
<p>Prioritisation looks like asking someone to cut a roadmap in half and defend the cut. Ownership looks like a take-home where the brief is deliberately incomplete and you watch what they do with the gaps. Systems looks like asking how they would redesign a process that currently takes ten people and three weeks. Taste looks like critique. Alignment looks like reference checks taken seriously. Insight looks like asking what they actually know about your customer that you do not.</p>
<p>Each of these is harder to fake than &quot;show me your prompts.&quot; That is the point.</p>
<h2>What this means for how to be hired</h2>
<p>Build evidence for the two or three letters that matter most for the role you want. Show the work. The CV that demonstrates two of these well is going to outperform the CV that lists ten AI tools every time. The roles that are hardest to fill in 2026 are not the ones that need the most tool fluency. They are the ones that need the most judgment.</p>
<p>If you are early in your career and worried that AI is closing doors, the doors that are closing are the ones that opened to people who ticked boxes. The doors opening are the ones that open to people who can demonstrate even one of these six convincingly. The bar is not higher because you need more credentials. It is higher because the credential everyone is reaching for is the wrong one.</p>
<h2>Closing</h2>
<p>AI-native, properly defined, is just this. Six traits, weighted by role. Recruiters who can hire for them will fill teams that compound. Recruiters who keep optimising for speed and tool fluency will fill teams with slop cannons.</p>
<p>The teams that win the next three years are the ones who built the rubric before the language caught up.</p>


The six traits recruiters should actually hire for in an AI-native world

POST-AI

<p class="lede">Three times in the last thirty years, the scarce thing shifted. Each time, a different type of person won. The people still running the previous game wondered what happened.</p>
<p>The third shift is happening now. Most people have not noticed.</p>
<h2>Three eras, three scarcities</h2>
<p><strong>Information.</strong> Before the internet, knowledge was scarce. Access was the moat. Institutions controlled the pipes. Credentials were the ticket in. Then information became free and that edge disappeared.</p>
<p><strong>Attention.</strong> Social media made content abundant. Anyone could publish. The new scarce thing was attention. The people who won were watchable. Follower counts. Virality. Aesthetic consistency. Being right mattered less than being compelling. Volume beat depth.</p>
<p><strong>Perspective.</strong> AI makes content and creativity abundant at scale. Not just searchable. Generatable. Anyone can produce a carousel, a caption, a thousand words, a UGC script. The volume of output in the world is increasing by an order of magnitude.</p>
<p>When everything is produceable at average quality, nothing at average quality is scarce.</p>
<p>What becomes scarce is a genuinely distinct lens. Not an opinion. A way of seeing.</p>
<h2>The bifurcation</h2>
<p>Both things will happen at the same time. Most people are only watching one of them.</p>
<p>The slop layer expands. More content, lower average quality, shorter average lifespan. Templates at scale. Hooks without argument. Engagement ticks. Nothing compounds. Nobody remembers it a week later. The social-era metrics still move, but they move toward nothing.</p>
<p>Underneath that, something different forms. Perspectives specific enough to find the people who needed exactly that lens. Not mass distribution. Precise distribution. AI-powered search and recommendation routing on specificity and relevance, not on follower counts or domain authority. A personal site with a genuinely original point of view outranking an optimised generic one. A niche take reaching its exact intended pocket faster than a broad take reaches anyone.</p>
<p>Two games, running in parallel. The slop game gets more crowded and less valuable simultaneously. The perspective game opens up.</p>
<div class="pull-quote">The social-era playbook confuses activity for compound interest. They are not the same thing.</div>

<h2>Why the social-era playbook fails here</h2>
<p>The instinct built in the attention era is to optimise for reach. Hooks. Posting frequency. Engagement loops. That instinct made sense when attention was the scarce resource. It is the wrong move now.</p>
<p>Perspective does not distribute like content. It compounds differently. A genuine take on a specific problem, written once, will find its intended audience through AI-routed search in a way that a well-formatted generic post never will.</p>
<p>The question is not how often you post. It is whether the perspective is worth having in the first place.</p>
<h2>What this means</h2>
<p>For the first time, original perspective has a structural distribution advantage over optimised reach. No institution required. No follower count. No budget. AI is the first distribution layer that actually routes on the quality of the thinking rather than the size of the audience behind it.</p>
<p>This is the opening most people are not seeing because they are looking at the slop layer and calling it the whole picture.</p>
<div class="pull-quote">The bottleneck moved entirely to the thought side. That is both the problem and the opening.</div>


The three eras of distribution, and why perspective is the only thing that compounds now

After attention

<p class="lede">The loudest voices telling you that you are behind on AI have something to sell. Urgency converts. Anxiety moves product. The &quot;you are already late&quot; narrative has one job and helping you think clearly is not it.</p>
<p>Being early is not the same as being good.</p>
<p>Social media proved this. The first people on TikTok did not build the most durable audiences. The first people on Twitter did not end up with the most influence. Being first meant more reps. Whether those reps compounded into something real depended on something else entirely. We remember the ones who were early and rose. We forget the much larger group who were just early.</p>
<p>Survivor bias is doing a lot of work in the AI discourse right now.</p>
<h2>What the foundation actually is</h2>
<p>The thing that determines whether AI makes you better is not how long you have used it. It is what you brought to it.</p>
<p>Critical thinking. Writing. Reasoning from first principles. The ability to read an output and know when it is wrong. Without these, AI does not amplify judgment. It amplifies confidence without judgment. Faster output, same quality of thinking underneath.</p>
<p>I graduated just before AI mattered enough to change how people worked. I built writing and reasoning under pressure before the tool existed to shortcut it. By the time I started using AI seriously, I had a foundation that could calibrate it. I could tell when it was wrong. I could push past the first draft. I could direct it rather than follow it.</p>
<p>That is not a disadvantage of timing. It is the thing that makes the tool actually work.</p>
<p>The person who grew up autocompleting their thinking has more reps. They may also have developed a dependency that is harder to see from the inside. The literature on overreliance is early but consistent. The debate happening around children and social media applies here at the adult level too. More exposure is not the same as better exposure.</p>
<div class="pull-quote">Without the right foundations, AI does not amplify judgment. It amplifies confidence without judgment.</div>

<h2>The efficiency trap</h2>
<p>The dominant AI metric is efficiency. Hours saved. Assets shipped. Cost per output. This is the wrong race.</p>
<p>It is dropshipping dressed as strategy. When ecommerce was new, the arbitrage was real. Then everyone got in. The edge closed. The moat was a timing window, not a durable advantage.</p>
<p>The efficiency gains from AI work the same way. Real, temporarily, and not a competitive moat. The person producing three times the assets of their competitor has an advantage until their competitor does the same. That takes months. Not years. What is left when the arbitrage closes is the quality of the thinking underneath the output. That was always the real variable.</p>
<p>The people building something durable are not optimising for output. They are redesigning how they think, how fast they learn, and how clearly they can see what is actually happening in their work. None of that shows up in hours saved.</p>
<h2>What intentional actually means</h2>
<p>Intentional AI is not a pace. It is a posture.</p>
<p>You can move slowly and be intentional. You can move fast and be reckless. Speed is not the variable. The question is whether you are thinking clearly about what is changing, and what it means for the decisions ahead.</p>
<p>Are you thinking about which parts of your work AI changes fundamentally, not just in speed? Are you building skills that compound, or substituting AI for the skill-building itself? Are you measuring the right things, or only the ones that are easy to count?</p>
<p>The person who has not touched an AI tool but is thinking clearly about the shift is better positioned than the person using it every day without asking why.</p>
<div class="pull-quote">Intentional AI is not a pace. It is a posture.</div>

<p>The question is not when you started. It is whether you are thinking clearly about what you are doing. That is a standard anyone can meet today.</p>


On intentionality, overreliance, and why being early to AI is not the same as being good at it

You are not behind

<p class="lede">This is written to designers, and only really for designers. Most of the noise about AI and design lands on either side of a tired axis. AI replaces designers. Or AI slop hollows out the discipline and craft is the moat. Both are arguments about the wrong layer.</p>
<p>The quieter story is the one worth your attention. The frictions that have shaped your craft for a decade are softening. Not all at once, not all yet, but the direction is clear and the rate of change is accelerating. I am writing this the way I would write it to a designer on my own team. Not to convince you of anything urgent. To make sure you see it before everyone else is.</p>
<h2>The frictions that defined the craft</h2>
<p>Design has run on a set of frictions that most designers treat as facts of life. They are not facts of life. They are artefacts of a workflow that is starting to break.</p>
<p>The mockup is a lie. Figma is static. Real sites have hover states, loading states, error states, animations, scroll behaviours, breakpoints. You ship one version. Engineering invents twenty more, often guessing wrong. You push back. The loop runs.</p>
<p>Spacing and type drift. You set a heading at 32px. Engineering writes 2rem. The unit is different, the rendering is subtly off, and the built site stops feeling like the design. Multiply that across hundreds of values and the design language goes soft.</p>
<p>The iteration asymmetry. Thirty seconds in Figma. Four hours in code. You iterate freely. Engineering pays the cost. Resentment compounds. Decisions slow.</p>
<p>The handoff. Briefly the celebrated artifact of the discipline. More recently, the thing that has quietly capped how much of your taste actually makes it to the built product.</p>
<p>These are not annoyances. They are the ceiling on your leverage as a designer for a decade. Strip them out and the ceiling moves.</p>
<h2>A compressed history</h2>
<p>The last ten years are easy to summarise. Sketch and Zeplin named the handoff. Figma killed Sketch with multiplayer. Webflow and Framer collapsed the entire pipeline for marketing sites. Then AI showed up in the loop. Each step closed a piece of the artifact-to-build gap. The current step is closing most of what is left.</p>
<h2>What no-code actually did</h2>
<p>The no-code wave was real but bounded. Webflow, Framer, Squarespace, Bubble. They collapsed the designer-to-developer pipeline for marketing sites, portfolios, simple e-commerce, internal dashboards. They got non-designers to 80% on those categories. Wrote the HTML and CSS for you. Wrapped the canvas.</p>
<p>They never got past 80% on serious product design. The ceiling was not skill-based. It was structural. The artifact lived inside a locked platform. You could edit the canvas, not the code. The moment you wanted something the platform did not anticipate, you fell off the cliff. Most product UI lives on the other side of that cliff.</p>
<p>The no-code chapter ended without taking the parts of the discipline that mattered most. The artifact was never real.</p>
<h2>What is structurally different now</h2>
<p>AI does not produce a locked artifact. It produces real code, in your codebase, in your conventions, that you can read, edit, ship, and version. That single shift moves the ceiling.</p>
<p>The moves of the last eighteen months, named without commentary. Figma launched Dev Mode properly. Code Connect maps a Figma component to the actual component in your codebase. Anthropic released MCP. Figma&#39;s MCP server lets Claude and Cursor read your Figma file directly. Claude generates working interfaces from a prompt. v0, Bolt, Lovable, Cursor. Each one on its own is a small step. Together they collapse the build step.</p>
<div class="pull-quote">The Figma file is no longer the source of truth. The codebase is.</div>

<p>That is the sentence to sit with.</p>
<h2>The personal proof</h2>
<p>I built my own website over a weekend. No design background. No Figma. No Tailwind training. No reference discipline. Wrote a sentence describing what I wanted. Claude generated the structure. I judged it live, prompted in language, iterated. Dropped into the CSS for the last few values. Shipped.</p>
<p>The site is fine. It is not good. The typography is slightly off. The spacing breathes too generously in some places, too tightly in others. The animation timing is not quite right. You can feel the missing taste.</p>
<p>That is the point.</p>
<p>If I can run the loop end to end on entry-level inputs and produce something that exists at all, a designer with taste, a real reference library, and basic CSS can produce something that actually lands. The loop is the same. The inputs are everything.</p>
<h2>The shape of the loop in practice</h2>
<p>Spec and references at minute one. Type system, colour palette, spacing scale, three reference sites. Treat it like a brief, not a prompt. This is the single largest force-multiplier in the whole workflow.</p>
<p>Optional: a Figma frame to give the AI a starting base. Useful for product UI where the structural decisions are heavy.</p>
<p>AI generates the structure in real code.</p>
<p>You judge the live output. Iterate in language. &quot;More breathing room around the heading.&quot; &quot;Warmer palette.&quot; &quot;Less corporate.&quot; &quot;Tighter density on the cards.&quot;</p>
<p>Optional: Claude or similar for prototyping animation. The thing that used to need a Lottie file or a developer&#39;s afternoon is now a conversation about easing curves and durations.</p>
<p>Final five percent in CSS. Exact values. Edge cases. The pieces the model keeps getting wrong.</p>
<p>Pair with a developer for performance, accessibility, and the weird Safari bug. That handoff still exists. It is much smaller and much more interesting than it used to be.</p>
<h2>What is not there yet</h2>
<p>I do not want to oversell this. The seams are real.</p>
<p>Real content still breaks layouts. The model does not know your data, your titles, your edge cases.</p>
<p>Animation timing is uneven. The model has opinions about easing and they are often wrong.</p>
<p>Accessibility edge cases still need a manual pass. Contrast on hover, focus indicators, reduced-motion fallbacks. The model improves. It does not yet replace the audit.</p>
<p>Bidirectional sync is messy. Figma to code is largely solved. Code back to Figma, when an engineer changes something, is still an open problem.</p>
<p>Context drift over long threads is real. The model forgets early decisions. Good operators restate the spec periodically or start fresh threads with a tightened brief.</p>
<p>The loop is closing. It is not closed.</p>
<h2>The skill stack that compounds</h2>
<p>Four inputs, in order of how much each one bends the curve.</p>
<ul>
<li><strong>Taste.</strong> The largest single input. Years to build. Untrainable in a weekend. The skill of looking at three options and knowing which is better, and being able to say why.</li>
<li><strong>Reference fluency.</strong> Knowing what good looks like across enough domains to pull a relevant anchor in seconds. Editorial sites, product sites, portfolios, niche aesthetics. Built through daily looking, not weekly.</li>
<li><strong>CSS vocabulary.</strong> A weekend. Padding, margin, line-height, letter-spacing, max-width, gap, flex, grid. Just enough to be precise when prompting and to read AI output without flinching.</li>
<li><strong>Spec discipline at minute one.</strong> A few projects to learn. The biggest force-multiplier in the loop and the one most designers still skip.</li>
</ul>
<p>Three and four are weekends. One and two are years.</p>
<div class="pull-quote">The leverage gap between the designer with AI and the non-designer with AI is not the loop. It is everything you bring to it.</div>

<p>That asymmetry is good news for designers. The shift you have been told to fear rewards the inputs you have already been building.</p>
<h2>How to prepare</h2>
<p>Concrete and short.</p>
<ul>
<li>Spend a weekend on CSS basics. Not to write it from scratch. To name it precisely when prompting and to read AI output without panic.</li>
<li>Build a reference library you actually use. Three sites for editorial, three for product, three for personal. Pinned somewhere you open daily.</li>
<li>Treat the first prompt of any AI design session like a real brief. Type, colour, spacing, references. A tight spec at minute one beats ten prompts of fixing.</li>
<li>Run one full project without Figma. Even badly. The point is to feel the new loop in your hands.</li>
<li>Pair with a developer for the last five percent. The handoff that survives is smaller, sharper, and more collaborative than the one you have been complaining about.</li>
</ul>
<h2>The close</h2>
<p>Your craft has been quietly clipped for a decade by the gap between the artifact and the build. That gap is closing. The designers who lean into the loop will be more leveraged, not less.</p>
<p>Treat this as a heads-up, not a warning. The shape is not all there yet. The direction is not in doubt.</p>


Dive in