The Taste Moat: How to Build the Skill AI Can’t Copy

The most repeated line in my feed this month is some version of “anyone can build anything now.” At the frontier labs it is already literal. Anthropic says AI writes somewhere between 70 and 90 percent of its code company-wide. Across the wider industry the rigorous number for production code sits closer to 27 percent, but the direction is not in dispute. The cost of producing a competent draft of almost anything, a landing page, a function, a brand, a pitch, is collapsing toward zero.

Here is the part nobody wants to sit with. If everyone can generate ten competent versions of a thing before lunch, competence stops being the edge. The scarce act is no longer making. It is choosing. Knowing which of the ten is right, and why, and being willing to throw out the other nine without flinching.

That skill has an old, slightly embarrassing name. Taste.

Most founders treat taste like weather. Something you have or you don’t, a gift handed out at birth to a lucky few who can pick the right font and the right feature and the right thing to say. I think that belief is wrong, and I think it is about to get expensive. Taste is not a gift. It is skilled intuition, the same kind a chess player or a firefighter builds, and the research on how skilled intuition forms is clear about the exact conditions that build it. Those conditions are things you can engineer on purpose.

I have shipped products that were technically clean and still felt wrong, and I have watched worse-built products win because they felt right. Every time, the gap was taste, and every time I could trace where mine had failed to develop. This is a field guide to building taste deliberately, written for founders and builders who just lost their old moat and need to know where the new one is.

Why competence stopped being a moat

For most of my career, being able to build the thing was the whole game. If you could write the code, design the screen, write the copy, and ship it before the other person, you won. Execution was scarce, so execution was the moat.

That assumption quietly died. When a model can produce a working first draft of nearly anything in seconds, the supply of competent output goes vertical. And the price of anything in infinite supply goes to zero. A competent React component, a competent cold email, a competent logo, a competent SQL query. None of those are scarce anymore. You can have as many as you want.

The thing that did not get cheaper is knowing which one is right. A model will happily give you ten landing pages, and nine of them are forgettable, and it cannot tell you which is the keeper because it has no stake in the outcome and no opinion about your customer. That judgment is the work now. One writer put it better than I can: refusal, knowing what to throw out and why, is the scarce act in a world where anyone can generate ten competent drafts before lunch.

You already pay for this skill without naming it. Think about the tools builders fight to use. Linear, Figma, Notion, Superhuman, Airtable. None of them won on a feature checklist. They won because they feel better to use, and “feels better” is just taste made physical. Someone on those teams could tell the difference between fine and right, again and again, across thousands of small decisions, and that accumulated judgment is the moat competitors cannot copy by reading a changelog. Steve Jobs built the most valuable company on earth partly on a refusal: he would kill good work because it was not the right work, and he could say why.

So the uncomfortable question for every founder right now is simple. If the model can build what you build, what exactly are you for? My answer is that you are for the judgment the model does not have. That judgment is trainable, and almost nobody is training it on purpose. That is the opening.

The Taste Gap: where value moved

Picture two curves over time, both moving as AI gets more capable. The first is the cost to produce a competent version of a thing. It is falling, fast, toward the floor. The second is the value of knowing what is actually worth producing. It is rising, because every cheap draft the world generates makes the act of choosing the right one more valuable, not less. Somewhere those curves cross. The space that opens up to the right of the crossing is what I call the Taste Gap, and it is where the new returns live.

Two things make this more than a pretty chart. First, the curves are not symmetric. The cost curve has a floor at zero and stops falling once it gets there. The value curve has no obvious ceiling, because there is no upper limit on how much it is worth to consistently pick the right thing in a market drowning in fine ones. Second, the gap compounds. Every founder who keeps shipping competent-but-wrong work widens the gap for the few who ship right work, because the wrong work trains customers to crave the right work even more.

Before going further, I need to define the word precisely, because most people use “taste” to mean something fuzzy and aesthetic. Taste is calibrated judgment about quality in service of a goal. It is not personal preference, and it is not which color you like. It is the ability to tell, quickly and correctly, whether a thing is good for the job it has to do, and to say why it is or is not. The “and say why” is load-bearing, and most of this guide is about it.

Taste is skilled intuition, not a gift

The belief that taste is innate is the most expensive idea in this whole conversation, because if you think you either have it or you don’t, you will never train it. You will outsource it instead, to an advisor, to a trend, to whatever the model spits out, and your own judgment will stay frozen at whatever level it reached by accident.

The research says the gift theory is wrong. In 2009, Daniel Kahneman and Gary Klein published a paper with a wonderful subtitle: “A Failure to Disagree.” Kahneman had spent a career showing that human intuition is often biased garbage. Klein had spent his showing that expert intuition is real, that a firefighter can feel a floor about to collapse and a nurse can sense an infection before the chart shows it. They sat down expecting a fight and instead agreed on something precise: when intuition is genuine skill, and when it is noise.

Their answer was two conditions. First, the environment has to be high-validity, meaning it contains real, repeating patterns that can in principle be learned. Second, the person has to have had prolonged practice in that environment with rapid, clear feedback. Meet both and intuition becomes a kind of compressed expertise: the chess master who sees the right move without calculating, the editor who knows a sentence is wrong before knowing why. Miss either condition and you get confident nonsense, the stock picker who feels certain and is no better than chance.

Taste is exactly this kind of skilled intuition, pointed at quality instead of chess. Which reframes the whole thing. People with great taste did not win a genetic lottery. They accumulated an unusual number of high-quality reps in an environment that gave them honest feedback about whether they were right, usually without calling it training. The designer raised on great design who shipped to real users and watched what worked. The writer who read obsessively and got edited hard. They built valid cues and got feedback, the two conditions, by luck or obsession.

And here is the trap hiding in the same idea. If you practice without the two conditions, you do not stay neutral. You build confident bad taste. You marinate in a low-quality input diet, you never get real feedback, and your judgment calcifies around whatever is popular and safe. Years pass and you call it experience. It is not. It is the same year of unexamined reps, repeated. Most people in most fields are here, which is the bad news and the opportunity in one sentence.

The Taste Engine: how taste is actually built

If taste is skilled intuition, then building it is not mysterious. You engineer the two conditions on purpose, over and over. I run it as a four-part cycle I call the Taste Engine. Each part exists to manufacture one piece of the Kahneman and Klein recipe, and the cycle is what makes it compound.

Input Diet. You cannot recognize great if you have only ever seen fine. The fastest way to upgrade taste is to upgrade what you look at, on purpose, every day. Stop scrolling the average and study the best work in your domain until the difference between great and good becomes obvious to you. This is how you learn the valid cues. Painters have done it for centuries by copying masters in the gallery. Your gallery is whatever you choose to put in front of your eyes.

Deliberate Reps. Consuming is not enough. You have to produce, and the goal of the producing is calibration, not output. Make the thing, set it next to the best example you have, and feel the gap. Most of my own taste jumps came from copying something I admired closely enough to understand why it worked, then trying to make my own and watching where mine went wrong.

Real Feedback. This is the condition almost everyone skips, and skipping it is why so many experienced people stopped improving a decade ago. Your reps need a high-validity environment that tells you the truth fast: real users, the market, a brutal and trusted critic, or a disciplined side-by-side against the best. Slow or flattering feedback trains the wrong thing. A loop that tells you the truth quickly is the single highest-return investment in your judgment, which is also the core idea behind loop engineering for AI agents: the verifier is the part that compounds.

Articulation. The last step is the one that separates good from great, and I gave it its own section below because it is that important. Once a thing feels right or wrong, force yourself to say why in words. “The hierarchy is fighting itself.” “This sentence is doing two jobs.” Naming it converts a one-time reaction into a rule you can apply tomorrow, teach to a teammate, and hand to a model. Skip it and your taste stays trapped in your gut, where it cannot scale.

The reason it is a cycle and not a checklist is the return arrow. Every turn raises your reference set, so your next batch of reps is judged against a higher bar, which produces sharper feedback, which forces finer articulation. That is the compounding, and it is why a year of real Taste Engine reps does not look anything like ten years of unexamined ones.

The first move, the Input Diet, is concrete enough to act on today. Here is how I audit mine.

The Taste Ladder: five levels and where you stall

Taste is not binary. You are not tasteless one day and tasteful the next. It climbs through levels, and naming the levels helps because most people stall at a specific one and never notice. Here is the ladder I use.

L1, Imitation. You copy what is popular. The clone of the trending app, the cold email swipe file, the design that looks like everyone else’s. It works sometimes, but you cannot say why, so you cannot adapt when the context changes. This is where AI output sits by default, and where you sit if you only prompt and paste.

L2, Recognition. You can tell good from bad on sight. You know the redesign is better, the sentence is sharper, the architecture is cleaner. But you cannot reliably make the better thing yourself. This is the critic’s level, and it feels like taste, which is exactly why it is a trap. The internet is full of people stuck at L2, confident and unable to ship.

L3, Production. You can actually produce good work, not just spot it. The catch is consistency. Some days it lands and some days it does not, and you cannot always explain the misses, which means you cannot fix them on purpose. Real, but not yet a moat.

L4, Articulation. You can produce good on demand and name the rules behind it. “This works because the hierarchy has one clear winner.” Now your taste is teachable, repeatable, and, crucially in 2026, delegable. You can hand the rules to a teammate or to a model and get work back that carries your judgment instead of the model’s defaults.

L5, Standard-setting. You define what good means in your domain, and other people calibrate against you. This is the level where taste becomes the moat I keep pointing at, the same one behind the products people describe as just feeling better. It is also one of the hard ceilings a one-person company runs into, because standard-setting is the work that does not compress no matter how good your tools get.

The wall that matters is between L2 and L3. Almost everyone can climb to recognition by consuming, because recognition only needs the Input Diet. Crossing into production needs the rest of the Taste Engine: the reps and, above all, the real feedback. The armchair critic skipped the reps. The reason AI makes this wall more dangerous is that a model will happily carry you at L1 forever, producing competent imitation, so you can feel productive while your own taste never leaves L2. The founders who pull ahead are the ones who use the cheap output as reps to climb, not as a crutch to stay still. That is also the heart of the cognitive debt founders take on when AI does the thinking.

Articulation: turning gut into rules

If you only take one practice from this whole piece, take this one. The jump from L3 to L4, from inconsistent good to reliable good, runs entirely through articulation: the habit of forcing your gut reaction into words.

Here is why it matters more than it sounds. A reaction you cannot name is trapped. It fires or it does not, you cannot improve it, you cannot teach it, and you certainly cannot hand it to an AI agent. The moment you name it, “the problem is that two elements are competing to be the most important thing,” it becomes a rule. Rules travel. You can apply the same rule next week, give it to a junior teammate, and put it in the instructions for the model so its output finally matches your standard instead of the average of the internet.

This is also the bridge between human taste and AI scale, and it is the most undervalued founder skill of this era. The reason your AI output looks generic is that you are feeding the model your gut, which it cannot read, instead of your articulated rules, which it can follow. People who can articulate taste get dramatically better work out of the same model than people who cannot, because they are programming it with judgment. That is the practical version of treating taste as a faculty you reserve and direct rather than delegate.

The practice itself is simple and slightly annoying, which is why few people do it. Every time something feels off, stop and ask one question that surfaces the rule. Then write the rule down. Over a few months you build a private file of rules that is, quite literally, your taste made portable. Here is what that looks like in motion.

Notice that none of those rules are about beauty. Each one ties a reaction to a goal: the screen’s job, the reader’s belief, the user’s moment. That is what separates taste from opinion. Opinion stops at “I like it.” Taste says “it is right for this job, and here is the rule that makes it right.”

Watch taste work: ten outputs, one decision

Theory is cheap, so here is taste operating on real AI output, the situation you are in every day now. The brief is “landing page for our new pricing.” You ask the model and it hands you ten competent versions in a minute. Every one is clean. Every one would have taken a designer a day in 2019. Now what?

The level-one move is to pick the prettiest and ship it. That is what the cheap-output world trains you to do, and it is exactly how you end up with a page that looks fine and converts like everything else that looks fine. The taste move starts with a rule from your file: every screen has one job. So the first question is not “which is prettiest” but “what is this page’s one job?” Say it is to make a hesitant visitor pick the middle plan without thinking too hard.

Now the ten outputs sort themselves fast. Six bury the middle plan in a three-column grid where every plan shouts equally, so they fail the one job and you kill them without guilt. Three make the middle plan visually dominant but explain too much, adding friction right where you wanted ease. One gets the hierarchy right and the friction low. That is your keeper, and you can say precisely why, which means you can defend it when a teammate prefers a different one for reasons that have nothing to do with the job.

Then comes the part the pure curators skip. None of the ten nailed the one line that removes the visitor’s real fear about being locked in. So you write it yourself, by hand, and drop it onto the keeper. That is the authorship that kept your taste sharp through the whole exercise, and it gives you a new rule for the file: name the buyer’s specific fear and answer it next to the price. This is the loop running in miniature, the same judgment-first habit behind a real founder operating system, and it took four minutes, not a day. Speed went up. Taste did the deciding.

Why taste is the one moat that compounds

Plenty of things look like moats and are not. A feature gets copied in a weekend. Your clever code is sitting in the model’s training data, which means a competitor can ask for the same thing. Even your best copy can be pasted into a prompt and remixed. Taste is different, and it is worth being precise about why, because the difference is the whole investment case.

First, it cannot be copied from your output. A competitor can see every choice you shipped and still not reproduce the judgment that chose them, because the judgment never appears in the artifact. It lives in the person doing the choosing. You can hand a rival your entire codebase and your roadmap, and if they do not have the taste to know which parts matter, they will rebuild the wrong things. This is exactly why so many AI startups are stuck, and why I keep warning founders about the wrapper trap of shipping commodity output. The model gave everyone the same output. Only taste tells you which version to keep.

Second, it appreciates instead of depreciating. Most assets lose value the moment you own them. Taste does the opposite. Every turn of the Taste Engine raises your reference set, which sharpens the next judgment, which compounds. A founder two years into deliberate practice is not twice as good as a founder one year in. The gap is wider than that, the same way interest is.

Third, it transfers. The person with real taste in product tends to develop taste in hiring, in writing, in deals, faster than a beginner would, because the underlying skill is the same one in every case: calibrated judgment about quality in service of a goal. So taste does not compound in one function. It compounds across the whole company, which is why a founder with it can be the final judge on the few decisions that actually decide the outcome, the kind of high-stakes calls I wrote about in the two-speed founder framework.

Fourth, at the top of the ladder it sets the standard. When you reach L5, you are not competing on the market’s definition of good. You are defining it, and the market calibrates to you. That is what the products people describe as just feeling better actually have. Not a secret feature. A team that could tell right from fine, thousands of times, until the difference became the brand. That is a moat you can build starting today and nobody can buy.

The contrarian take: taste without authorship is a trap

Here is where most people are about to take the wrong lesson, and it is the seductive one. If execution is free and taste is selection, then the obvious move is to become a pure curator. Let the model produce, and spend your time picking the good one. Why make anything yourself when judging is the high-value act?

Because taste built only on selection rots, and it rots quietly. Remember where taste came from: reps and feedback, the act of making things and being told the truth about them. That is authorship. The moment you stop making and only select from a machine’s output, you stop paying the cost that built your judgment in the first place. The Taste Engine stalls. Your reference set stops rising. And within a few months your taste drifts down to L2, then drifts toward whatever the model nudges you to like, because you no longer have an independent standard to judge against.

There is a sharper version of the risk. If you only review machine output, you slowly become a reviewer of a machine-led process with no real stake in it, and reviewers with no stake converge on the defaults. You start approving the competent-but-generic thing because you have lost the reps that would have told you it was wrong. Taste without authorship is fragile. It feels like an edge while it is quietly decaying, which is the most dangerous kind of decline, the kind you do not notice until a competitor with real taste makes your work look exactly as average as it became.

This does not mean do everything by hand. That would be its own mistake, and a slow one. It means keep your hands in the specific work that trains your judgment, and automate the rest aggressively. Use AI to run more reps, not to skip them. Generate ten versions, yes, but then make the eleventh yourself, compare, and articulate the difference. Selection and creation are not separable skills. The selection only stays sharp because you are still creating.

What to do Monday morning

None of this matters as theory. It is a practice or it is nothing. Here is the smallest version that actually moves your taste, starting this week.

1. Make one input-diet swap. Pick a single domain from the table above and make this week’s one swap. One hour of studying the best replaces one hour of scrolling the average. That is the entire change, and it is the highest-return one.

2. Start a taste file today. Open one document. Every time something feels off or feels great, stop, ask the question that surfaces the rule, and write the rule down. Three entries this week. In a year this file is your judgment, made portable and teachable.

3. Build one honest feedback loop. Take one thing you are making and put it in front of reality fast this week: real users, a critic who will not flatter you, or a disciplined side-by-side against the best example you can find. Slow or kind feedback does not count. You are engineering the high-validity condition on purpose.

4. Run one rebuild rep. Once a week, take a piece of work you admire and rebuild it by hand. Not to ship it, to feel why it works. One hour. This is the single fastest taste accelerator I know.

5. Find your stall level and treat the right disease. Be honest about where you are. If you can judge but not produce, you are at L2 and your problem is reps and feedback, not more consuming. If you produce well but inconsistently, you are at L3 and your problem is articulation, so the taste file is your move. Diagnosing the level stops you from taking the wrong medicine.

6. Program your AI with rules, not vibes. Take five rules from your taste file and put them directly in your model’s instructions. Compare the output before and after. This is the practical bridge between human judgment and machine scale, and it is why people with articulated taste get work out of the same model that everyone else cannot.

Do this for ninety days and your taste moves a full rung on the ladder. In a world where everyone can produce infinite competent output, moving one rung in judgment is not a small edge. It is the only edge left that compounds, and it is the one nobody can take from you.

FAQ

What does taste actually mean for a founder?

Taste is calibrated judgment about quality in service of a goal. It is not aesthetics and it is not personal preference. It is the ability to tell, quickly and correctly, whether a thing is good for the job it has to do, and to say why. The “say why” part is what separates taste from a vague opinion, and it is the part you can train.

Is taste innate, or can it be learned?

It is learned. Taste is a form of skilled intuition, the same kind Kahneman and Klein studied in firefighters and chess masters. Their research found that genuine expert intuition forms under two conditions: a high-validity environment with real, repeating patterns, and prolonged practice with rapid, clear feedback. People with great taste are not gifted. They got more high-quality reps with honest feedback, usually without naming it as practice.

Why is taste called the new moat in the AI age?

Because AI is driving the cost of producing competent output toward zero. When anyone can generate ten competent versions of a thing in minutes, making is no longer scarce and choosing the right one is. That judgment cannot be copied from your output, because it lives in the person choosing, not in the artifact. So taste becomes the defensible edge precisely as everything else gets commoditized.

How do I build taste deliberately?

Run the Taste Engine: a four-part cycle of Input Diet (study the best, not the average), Deliberate Reps (produce to calibrate, not just to ship), Real Feedback (a fast, honest, high-validity loop), and Articulation (name why good is good and turn it into a rule). Each turn raises your reference set, so the next rep is judged against a higher bar. That is what makes it compound.

What is the difference between recognizing good work and actually having taste?

Recognition is level two on the Taste Ladder. You can spot good from bad, but you cannot reliably produce good yourself. That is the critic’s level, and most people stall there because consuming alone gets you to it. Real taste, the kind that is a moat, is production and articulation, levels three through five, and crossing into them requires reps and real feedback, not more scrolling.

Can AI have taste, or replace my taste?

No, and the reason is structural. A model has no goal of its own and no stake in your outcome, so by default it produces competent imitation, level one on the ladder. It can execute taste you have articulated into rules, which is powerful, but it cannot supply the judgment about what is worth making. Used well, AI is a way to run more reps. Used as a crutch, it keeps your own taste frozen while you feel productive.

How long does it take to improve taste?

It depends far more on the quality of your feedback than on raw time. A year of unexamined reps in a low-feedback environment moves your taste almost nowhere, which is why some people have twenty years of experience that is really one year repeated. Ninety days of deliberate practice with a real feedback loop will move you a full rung on the ladder. The variable is honesty of feedback, not hours.

How do I get better output from AI by using my taste?

Articulate your judgment into explicit rules and put them in the model’s instructions. The reason most AI output looks generic is that people feed the model their gut, which it cannot read, instead of their rules, which it can follow. Pull five rules from your taste file, add them to the system instructions, and compare the output before and after. That is how human judgment turns into machine scale.