How to actually work with an AI assistant

The stack matters less than the method. This is what changed the output most, in rough order of impact.

Make it interrogate you first

The single highest-leverage habit. Before it writes anything, have it ask you multiple-choice questions about the ambiguous parts.

"Before you write any code, ask me questions one at a time about anything ambiguous. Give me options with a recommendation."

Almost all bad output traces back to a vague request rather than a weak model. Ten seconds of questions routinely saves an hour of confidently wrong work.

Say what you want, not how to build it

Describe the outcome and the constraints. Implementation detail you volunteer is detail it will follow even when it is wrong.

• Weak: "add a useEffect that fetches bookings and sets state"
• Strong: "the host dashboard should show this week's bookings, newest first, and stay fast with a few hundred rows"

Dictate instead of typing

Voice-to-text is roughly three times faster, and speed matters here for a non-obvious reason: you include context you would not have bothered to type. The background detail is what makes the answer good. You will fix the occasional mangled word — worth it.

Tell it to flag suspected mis-hearings rather than guessing at them.

Write the rules down once

Keep a rules file in the repo (CLAUDE.md) that the assistant reads every session. Every hard-won lesson goes in it. Otherwise you re-teach the same lesson monthly and it silently regresses in between.

Make it prove things instead of asserting them

"It works" is not evidence. Ask for the command output, the HTTP status, the screenshot. Assistants are strongly inclined to report success — a request for proof costs one line and catches real failures.

The corollary: give it the ability to check. Connect it to your database, your error tracker, your deploy logs. An assistant that can read your live systems is dramatically better than one guessing from the code alone.

Push back, and expect push-back

Ask it to argue against your plan before executing it. An assistant that agrees with everything is a very expensive autocomplete.

The part that will actually cost you: the loop

An AI pair removes the friction that used to make you stop. There is no "tomorrow" when the next feature takes four minutes. And it will offer you a next task at the end of nearly every response.

That suggestion is genuinely useful and it is also a slot machine. Unchecked, you spend a month building things nobody asked for while the question you needed to answer — does anyone want this — goes untouched.

On the reference build: 976 commits in one month, peaking at 95 on a national holiday. The result was a marketplace with ten directory categories, five dashboards, and two storefronts, while the core business earned tens of dollars a month. Almost none of that surface area was the constraint.

What works:

• Tell it explicitly: do not propose follow-up work unless I ask.
• Decide the week's scope before opening the editor, not during.
• Put a hard stop on the day and keep it.
• When it suggests the next shiny thing at the end of a response, that is the moment to close the laptop — not the moment to keep going.

Ship in reviewable pieces

One task, one branch, one pull request. Not process theatre: it is the only way to undo exactly one thing when something breaks two weeks later.