Sanity Check

Passing tests is necessary but not sufficient. Before declaring done, verify your code matches the real world.

When to Run

After implementation is complete and tests pass. Before committing or declaring done.

Checks

1. Run with Real Data

• Execute the program with real inputs (dry-run if available).
• Check output makes sense — wrong numbers, weird dates, implausible results → investigate.
• If no dry-run mode exists, call real endpoints in a throwaway script and confirm the response matches your code's assumptions.

2. Cross-Check External APIs

• Verify every field name, enum value, and response shape against actual API docs.
• Don't trust what you "know" — open the docs and confirm.
• If you can't verify a field, response shape, or identifier from docs or real responses, stop and ask. A wrong guess that passes tests is worse than a blocker that gets flagged.

3. Smell Test

• Does the output look right to a human?
• Are the numbers in the right ballpark?
• Would you trust this if you saw it in production?

4. Contract Boundaries

• Changed a public API/interface? Verify all callers still work.
• Touched persistence/state? Verify read/write roundtrip with real data.
• Changed config? Verify the program starts with the new config.
• Schema/state changes must be forward-compatible. Document rollback path.

If Something Smells Off

Don't rationalize it. Investigate. Don't burn tokens on workarounds — surface the issue.

Output

After running checks, report:

• Verified: what you checked and how
• Gaps: things you couldn't verify and why
• Confidence: ship it / needs review / blocked