patrickz
Before you start
A coding agent, a small repository and version control. Work on a branch or disposable practice copy.
Why this lesson exists
The archived posts behind “Give Your AI a Code-Review Conscience” treat prompts, project rules and tests as part of the product. A coding agent’s confidence is not a substitute for inspecting the change.
The idea this guide practices: Build an app with an AI agent, one proven slice at a time · I had a flight to Thailand. So AI rebuilt Mario with me in it. Prefer a story? Read the comic The Scope Creep Mecha Is My Fault.
Do the exercise
Save a definition of done
Open the project instructions file your coding agent reads. The name depends on the tool; AGENTS.md is one common example. Paste the checklist from the prompt below, delete any line that does not apply to this project, and commit the file so every future session starts from the same rules.
State the behavior
Describe a concrete before-and-after case. Identify the user action, expected response and boundaries. Ask the agent to inspect existing conventions before editing.
Review the proposed change
Request a file-level explanation and the reason for each change. Watch for unrelated rewrites, new dependencies and credentials in client code.
Test the failure that matters
Choose a test that would fail on the original bug or missing behavior. Include one ordinary case and one boundary. Run the project’s required checks.
Inspect and checkpoint
Read the diff, use the interface and record what was verified. Commit the scoped change so you can revert it. Keep untested behavior explicit. Before you stop, have the agent stop building and report pass, fail or not verified for each checklist item.
A prompt to adapt
Replace the bracketed parts with your own details.
DEFINITION OF DONE (save this in the project instructions file) 1. The main workflow works end to end, and you ran it. 2. Data survives a reload. 3. Errors show a message a person can understand. 4. Secrets stay out of browser code and out of the repository. 5. Each user sees only their own data (if the app has accounts). 6. A new user can open it without help. 7. Data can be exported. 8. The diff touches only files this task needs; ask before adding a dependency. 9. Report the commands you actually ran with their output, and list anything unverified. TASK Read the definition of done in [instructions file] first. Implement [specific behavior] in this repository, following existing conventions. Keep the change focused. Explain the diff, add or run a test that demonstrates the behavior, then check the project against every item and report pass, fail or not verified for each.
What this looks like
A test that merely checks a button exists does not prove that saving survives reload. Test the user-visible outcome rather than mirroring the implementation.
Check your result
Use evidence from your output. A confident explanation from the AI is not enough.
If it isn’t working
If a test passes without exercising the changed behavior, strengthen the assertion. If the agent reports tests it did not run, request actual output and inspect it.
Optional. Progress stays in this browser.
Where this came from
Adapted from the archived LinkedIn theme “Give Your AI a Code-Review Conscience.” See the post coverage and editorial method.
Prepared September 2026. Tools and interfaces change; use current official setup instructions. Session lengths are estimates.