AI is for the plumbing, not the judgment

A model can generate your entire product now. That is the easy part, and it is the part everybody is doing. The hard part is the decisions — what to build, what to cut, what the bar is — and those have not gotten easier. They got harder, because there is no slowness left to hide them behind. I have written before that the hardest part of building with AI is not the building; the practical version of that idea is a simple split: hand the model the plumbing, and keep the judgment yourself.

What does "plumbing" mean in an AI workflow?

Plumbing is any work where you know it is done when you see it. A checkout flow that works. A page that matches your tokens. A refactor that passes the tests. A summary that is accurate. These tasks have a spec, even if the spec only lives in your head, and the model is excellent at them — often better than you, and certainly faster.

The tell is the outcome. If a checkable result exists, the task is plumbing. Hand it to the model without guilt. That is not the part of the work you are paid for anymore.

What does "judgment" mean in an AI workflow?

Judgment is the work with no checkable outcome. What this product should be for, exactly. Which user to ignore. Whether the name is right. Whether the hierarchy of a screen tells the truth. Whether the copy sounds like a person with a point of view or like a committee.

There is no spec for any of that. The model cannot hold a standard it does not share, and it has no stake in your product. Ask it "is this good?" and it will tell you yes — that is what it is trained to do. The only person in the loop who can say no is you.

What does the data actually say about the split?

This is not a vibe. Anthropic's own Economic Index survey data shows that experienced Claude users name judgment and management as the capabilities the models still lack — the more experience a user had, the more likely they were to say it. The people closest to the tools are telling you exactly where the tools stop.

The business data agrees. Stripe's analysis of solo-founded startups found top-decile founders retained first-month customers at six times the rate of the median, and that AI-native solo startups pulled in almost double the revenue by month 24. The winners did not replace themselves with models. They used them to move faster and kept the calls — and retention is a direct measure of whether those calls were right.

How do you keep judgment when you work alone?

Three habits.

Name the lane before you prompt. Say it out loud: this is plumbing, or this is judgment. If it is judgment, the model is a second opinion, never the decision-maker.

Gate everything that ships. First-pass output is a hypothesis. Read it against your own standard and change what fails. This is the review pass I wrote about in your eval set is your taste, made measurable — the machine reads faster, but only you can say whether it is right.

Keep a slice by hand. Writing, editing, or a small piece of code done deliberately keeps your sense of "good" calibrated. Taste atrophies from disuse, and Paul Graham's Taste for Makers makes the case better than I can: the recipe for great work is very exacting taste plus the ability to gratify it. AI is the ability to gratify, at scale. The exacting taste still has to be yours.

What is the failure mode to watch for?

The losing move is delegation — letting the model make the decisions and then shipping the output because it looked finished. A team that uses AI badly just makes bad work faster, and a solo founder who outsources the calls becomes a very fast producer of someone else's average.

When you catch yourself asking a model "what should I build?" or accepting a first draft as final because it is fast, you have crossed the line. The answer is not to build slower. It is to be deliberate about what the model is allowed to decide. It decides how. You decide what and why.

The plumbing-versus-judgment split is not a productivity trick. It is the whole strategy. The models get better every quarter, and your job is to stay ahead of them at the one thing they are not getting better at — caring about your product more than they ever will. Use the machine for the work. Keep the calls for yourself.