Audit code that may have been generated by agents or rushed through an AI-assisted implementation pass.
This skill does not replace review, review-and-ship, proof-orchestrator, or aegis-structured-execution. It adds a sharper anti-generated-code lens.
Use When
Use when Vitor asks to:
- review AI-generated code;
- check for slop;
- clean up a Codex-generated implementation;
- audit agent work before commit;
- find speculative abstractions or fake safety;
- compare implementation against repo style after an AI pass.
Required Context
Inspect:
- changed files;
- surrounding source;
- repo instructions;
- tests and proof surfaces;
- original spec or user request when available.
Do not review only the patch if surrounding architecture determines whether the change is good.
Audit Lens
Look for:
- duplicated helper logic;
- generic wrappers that do not carry domain value;
- speculative options, flags, abstractions, or extension points;
- defensive fallbacks that hide real failures;
- broad try/catch blocks with weak error truth;
- stringly typed contracts where the repo uses types or schemas;
- fake TODOs, placeholders, or soft "future" language;
- code comments that narrate obvious behavior;
- tests that only prove mocks;
- proof claims not supported by fresh evidence;
- behavior that satisfies the prompt but violates repo conventions.
Output
Lead with findings.
## Findings
| Severity | File | Issue | Why it matters | Fix |
| --- | --- | --- | --- | --- |
## Keep
- Strong choices that should not be rewritten.
## Proof Needed
- Test, runtime, or review proof required before shipping.
## Suggested Patch Shape
- Smallest safe cleanup sequence.
Guardrails
- Do not rewrite code just because it looks generated.
- Do not flatten useful abstractions.
- Do not invent a new architecture during review.
- Do not claim code is AI-generated based on style alone.
- Do not treat style issues as blockers unless they create real maintenance, correctness, security, or proof risk.