ENG-08 · SEC. 01 Engineering & Coding

Hostile Reviewer Pass

Make the agent tear apart its own diff like a hostile senior reviewer before you ship it.

FORMAT: prompt
DIFFICULTY: beginner
TIME: 10 min
TOOLS: claude-code · cursor · codex-cli · gemini-cli
MODELS: any
COPIES: 0 so far

When to use this

The change is written, tests pass, and the agent is telling you it's done. Run this before you commit or open a PR, while the diff is still cheap to fix. It works best on diffs big enough to hide something: 3+ files, or anything touching auth, money, or data writes.

The pattern

Pastes as plain text

Review the current uncommitted diff as a hostile senior reviewer who did not write this code and gets no credit for being nice.

1. Run git diff and git diff --staged to get the full change. Review the diff plus enough surrounding code to judge it in context.
2. Hunt specifically for: broken edge cases, error paths that swallow failures, security holes (injection, missing authz, secrets in code), changes that contradict this codebase's existing conventions, and anything the diff changes but no test covers.
3. Report each finding with a severity:
   - BLOCKER: would cause a bug, data loss, or security issue in production.
   - SHOULD-FIX: works today but will bite someone later.
   - NIT: style or naming. Mention once, don't pad the list.
4. Do not edit, stage, or commit as part of this review. For every BLOCKER and SHOULD-FIX item, propose the smallest patch and affected test; wait for my explicit approval before implementing any selected fix.
5. If you find zero BLOCKERs, do not just say "looks good." Name the three riskiest spots you checked and what convinced you each one is safe.

Real example output

Reviewed diff: 4 files, +182/-31 (Stripe webhook handler for payment events).

BLOCKER (2):
1. routes/webhooks.ts:47: handler returns 200 before awaiting markInvoicePaid(). If that write fails, Stripe never retries and the invoice stays unpaid forever. Proposed patch: await the write and move the 200 after it; affected test: failed-write retry behavior.
2. routes/webhooks.ts:12: signature check reads req.body after the JSON middleware already parsed it, so raw-body verification always fails open with the prod config. Proposed patch: mount express.raw() on this route only; affected test: invalid-signature rejection.

SHOULD-FIX (1):
- No test for duplicate events (Stripe retries send the same event ID twice). Proposed test: second delivery is a no-op.

NIT: handlePaymentIntentSucceeded is 80 lines and could split. Left alone.

No changes made. Awaiting approval to implement any selected fixes.

Why it works

An agent grading its own fresh work is the most generous reviewer alive. Recasting it as a hostile outsider flips the incentive: the persona gets credit for findings, not for finishing. The zero-BLOCKER clause closes the last loophole, because "here's what I checked and why it's safe" is work, while "looks good" is free.

Related patterns

DBG-04Adversarial Fix Verification LoopHave a second, fresh-context pass try to prove the first agent's fix is wrong before you trust it.WRT-13The Skeptical Reader GauntletLoop hostile-reader objections against your draft, patch the real ones, repeat until it survives a clean pass.OPS-10Pre-Push, Pre-Merge Approval GateStop an agent from pushing or merging without your explicit go-ahead, every time.