Taste: What AI Cannot Do - Playbook

The thing the AI does not have

Describe ten possible features and the AI will help you build any of them, thoughtfully and well. It will not tell you which of the ten matters. It has no stake, no market, no memory of the user who called you frustrated last week. It optimizes for the task you hand it, not for whether that task was the right one.

So the most important questions in the whole project are the ones the AI cannot answer for you:

Is the thing the user asked for the real problem, or just what they think will solve it?
Of the four reasonable next builds, which one ships first, and which one never ships at all?
Is this feature good enough to ship, or am I polishing something nobody needs?
Are we even building the right product?

Those are taste questions. They are yours, always.

What taste decided on this project

The record is clear about which decisions were the human's, because the human's decisions are the ones that changed the direction of the whole thing:

The pivot. The product started as a general CRM for solopreneurs and became "the operating system for the trades." No amount of code generation produces that move. It came from looking at who was actually using it and what they were underserved on. It is the single highest-leverage decision in the history of the product, and it was pure judgment.
What to build in the photo-and-voice line. Faced with three plausible directions, the choice was the one that formed a revenue chain (walk the job, get a recap, get a quote, get paid), not the one that was the most impressive demo. That is a strategic call about where the money is, not a technical one.
The name. Renaming the product across the entire codebase was a one-day mechanical task for the AI and a months-in-the-making decision for the founder.
Pricing. Moving the marquee feature down a tier because locking it behind a higher price was, in the founder's words, "a churn machine." The AI can implement any pricing. It cannot feel which one loses you customers.

In every case the AI did the building and the human did the deciding, and the deciding is what made it a product instead of a pile of features.

Knowing when to stop

There is a specific failure the AI will lead you into if you let it: it will keep improving. Ask for refinement and it refines, forever, happily, past the point of any return. The instinct to say "this is good enough, ship it, leave the rest for next week" is a founder instinct, not an engineering one, and it has to come from you. Shipping is a taste decision. So is not building something at all. The restraint to leave the tenth idea unbuilt is as important as the judgment to build the first.

Taste is also knowing what is true

There is a second half to taste that is less about strategy and more about evidence. When something breaks, the AI will reason confidently from the code and hand you a plausible wrong answer, and then another one. The judgment to stop guessing and go read the actual evidence (the error trace, the log, the real network response, the live data) is the same muscle as knowing what to build: it is the refusal to accept a confident story in place of the truth. Reason from the ground, not from the narrative. (The verification module is this habit applied to bugs; here it is the general principle.)

Where taste comes from

Taste is not innate and it is not guessing. On this project it came from talking to actual users, roofers and realtors and contractors, in person and on the phone. The in-person tests did not just catch bugs; they reshaped the product. You cannot develop taste about a market from inside your editor. The judgment is yours, but it is informed by contact with the people you are building for, and that contact is itself work you have to go do.

The honest caveat: taste can be wrong

This is not a story about a founder whose every instinct was right. The marquee feature of this product, the one the marketing leads with, has near-zero adoption so far. That was a taste bet, and the bet has not paid off yet. Good taste is not being right every time. It is making the call with the best information you have, watching what actually happens, and being honest enough to see when a bet is not landing so you can change it. Taste without feedback is just ego. The decision log and the usage numbers are how you keep yourself honest.

The shape of the whole thing

Here is the through-line of the entire course. The AI collapsed the cost of building to almost nothing. That did not make judgment less important. It made judgment the entire job. When writing the code is free, the only things left that matter are deciding what to build, governing what ships, verifying it works, keeping it honest, getting it found, and knowing when to stop. Those are all taste. The leverage is real, and it concentrates every ounce of your value into the choices the machine cannot make for you.

Exercise

Keep a decision log. Write down each meaningful decision and why you made it. The forty session logs behind this course are exactly that, and they are why its story is recoverable at all.
Justify the build before you build it. For your next feature, write one sentence on why this one and not the other nine. If you cannot, you have not decided yet.
Practice killing. Find a feature you built that users do not use. Decide, honestly, whether to fix it or retire it. The willingness to retire your own work is the clearest sign you have taste and not just attachment.

Taste: What the AI Cannot Do