Collaboration Playbook
The workflow tells you which steps to take; this page tells you how to work with the AI — how to instruct it, review its output, step in, and recover when stuck. Make these habits and AI becomes a teammate worth having.
Back to the AI coding workflow →Direct
How to hand off work so AI gets it right.
Hand off one thing at a time
Why：Juggling several things at once, AI drops some — and you can’t tell which part broke.
Do this
“Just build the login page for now, leave the rest.” Then hand off the next piece.
Not this
“Do login, payments, and the admin panel all at once.”
Spell out the context and constraints
Why：AI can’t see what’s in your head; without context it just guesses.
Do this
Give the target users, the services to integrate, and what’s off-limits — then let it start.
Not this
Just say “make a shopping cart” and expect it to read your mind.
Define what “done” looks like up front
Why：With no definition of done, AI calls it finished while you think it’s only half-built.
Do this
“Done = can register, can log in, wrong password shows an error.”
Not this
Open-ended instructions like “just get it roughly working.”
Review
How to check AI’s output before you trust it.
Don’t accept what you can’t read
Why：Accepting code you don’t understand quietly signs you up for future maintenance pain.
Do this
Have AI explain it piece by piece until you can restate what it does in your own words.
Not this
It runs, so paste the whole thing and move on.
Check the output against your spec
Why：Code that runs isn’t proof it does the thing you actually asked for.
Do this
Map what it built back to your feature list, checking nothing’s missing or extra.
Not this
Seeing the screen work and calling it done, never circling back to the spec.
Intervene
When to step in and pull it back on track.
Stop the loop when AI spins
Why：When it circles the same error, pushing harder just makes it loop longer and messier.
Do this
Call a halt, restate the problem, or describe it from a different angle.
Not this
Just replying “still wrong, try again” and hoping it breaks out on its own.
Make it prove “done”
Why：“Should work now” isn’t evidence; an unverified “done” is often half-finished.
Do this
“Give me steps or a test I can run to prove it actually works.”
Not this
Hearing “it’s done” and shipping or delivering right away.
Recover
When it goes wrong, how to get back on course.
Shrink the task when it’s stuck
Why：Most stalls come from one step trying to do too much; shrinking it usually unblocks instantly.
Do this
Break the big task into two or three smaller steps that each finish on their own.
Not this
Re-explaining the same big task over and over, expecting a different result.
Start a fresh chat when it’s truly off
Why：When the chat gets long and context rots, AI clings to old wrong assumptions.
Do this
Open a new chat and restart with just the current code and a clear goal.
Not this
Endlessly patching inside an already-tangled conversation.