ENG-03 · SEC. 01 Engineering & Coding

Self-Verifying Build Loop

Make the agent build, run, diagnose, and retry until there's real passing evidence, not a hunch.

FORMAT: loop
DIFFICULTY: intermediate
TIME: 15 min
TOOLS: claude-code · cursor · codex-cli
MODELS: any
COPIES: 0 so far

When to use this

You're asking an agent to write or change code and you don't want it to declare victory after a change that merely "looks right." Use this whenever there's a build, test suite, or reproducible failing case to check against.

The pattern

Pastes as plain text

Implement this change and do not stop until you have real passing evidence.

CHANGE: [WHAT TO IMPLEMENT OR FIX]
BUILD/TEST COMMAND: [THE COMMAND THAT PROVES IT WORKS]

Loop:
1. Write or edit the code for the change.
2. Run BUILD/TEST COMMAND. Paste the actual output, not a summary.
3. If it fails: identify the root cause from the actual error message (not a guess). State the root cause in one sentence before touching code again.
4. Fix only what the root cause requires. Do not make unrelated changes.
5. Re-run BUILD/TEST COMMAND.
6. Repeat steps 3-5 until the command passes cleanly, or until you've tried 5 times, in which case stop and report what you tried and what's still failing.

Never report the change as done without pasting the final passing output. "Looks correct" is not evidence.

Real example output

Attempt 1: npm run build fails.
  TS2345: Argument of type 'string | undefined' is not assignable to type 'string' in src/lib/parseConfig.ts:42
  Root cause: config.env.API_KEY can be undefined at build time; function signature assumes it's always defined.
  Fix: added a guard that throws a clear error if API_KEY is missing, instead of loosening the type.

Attempt 2: npm run build
  > tsc --noEmit
  > vite build
  ✓ 412 modules transformed.
  build finished in 3.1s
  PASS

Reporting done: build is green, output above. No unrelated files touched.

Why it works

Agents left unchecked will often patch the symptom (silence a type error, add a try/catch) instead of the cause. Requiring a one-sentence root cause before each fix, plus pasted real output, blocks both premature "done" claims and cosmetic patches that don't actually fix anything.