• English

      • Review AI Agent output

      When your AI Agent says it is done, walk through the requirement. A visible page is only the beginning: verify persistence, restrictions, and whether the application builds.

      Use it as a user

      The routine in your first feature, creating, refreshing, editing, and then trying invalid input, is this kind of check: the normal path must work, data must actually be saved, and limits must hold when input is wrong.

      For permission requirements, switch accounts. Administrator access does not prove ordinary users have the correct data scope, and hidden buttons do not replace API authorization.

      Review the scope of changes

      Ask your AI Agent to list new and changed files. If the project already uses Git:

      git status --short
git diff

      git diff does not include untracked new files, so inspect both. Confirm that pages, APIs, and database changes fit together, unrelated features remain intact, and configuration secrets are not included in commits.

      For migrations, check that existing data is preserved. Do not rebuild a database just to fix a field. Correct merged migrations through new migrations; see schema changes.

      Run the project's checks

      Inspect scripts in the application's package.json. The default template provides:

      pnpm typecheck
pnpm test
pnpm lint
pnpm build
      CheckWhat it can find
      typecheckType, import, and API declaration mismatches
      testFailures in behaviors covered by existing tests
      lintProject-rule violations and some code issues
      buildClient/server build failures and missing deployment dependencies

      pnpm check runs lint, formatting, types, tests, and the build in sequence. A successful command does not prove all business behavior: pnpm test may succeed when no test files exist. Still verify actual interactions.

      For example, a page may work during development while server dependencies are listed only in devDependencies, causing deployment to fail. Build checks help catch this problem.

      Return specific feedback

      Describe the action, expectation, and actual result:

      I created ORD-001 and saw it in the list after saving.
It disappeared after refreshing the browser.
I expected it to remain.
Please check whether the page calls the save API and whether the API writes to the database.
After fixing it, verify creation, refresh, and editing, and report the actual results.

      For terminal failures, include the working directory, command, and full error. Exclude secrets, real customer information, and credential-bearing configuration.

      If your AI Agent proposes removing a check, forcing a type conversion, or resetting the database, ask for the reason and impact. Fix the cause and recheck existing behavior.

      Keep a record of the round

      Record verified behavior and remaining issues before continuing. See complex features for a step-by-step approach.