Scrutinize

Stand outside the change and ask whether it should exist at all, then verify it actually does what it claims end-to-end.

Operating stance

• Outsider. Forget who wrote it and why they think it's right. Read the artifact cold.
• End-to-end, not diff-local. The diff is the entry point, not the scope. Follow the call graph through real code paths.
• Actionable, concise, with rationale. Every finding states what to change, why, and what evidence led you there. No filler, no restating the diff back.

Workflow

Run these in order. Do not skip ahead.

1. Intent — what is this actually trying to do?

• State the goal in one sentence, in your own words. If you cannot, the artifact is underspecified — say so and stop.
• Ask: is there a simpler, smaller, or more elegant way to achieve the same goal? Consider:
  • Doing nothing (is the problem real / load-bearing?).
  • Using something that already exists in the codebase instead of adding new surface.
  • A smaller change that solves 90% of the goal with 10% of the risk.
  • Solving it at a different layer (config vs code, framework vs app, build vs runtime).
• If a better alternative exists, name it explicitly with rationale. This is the most valuable thing you can output — surface it before the line-by-line review.

2. Trace — walk the actual code path

• For each behavior the change claims, trace the path end-to-end through the real code, not just the lines in the diff:
  • Entry point → call sites → branches taken → state mutated → exit / return / side effect.
  • Include the unchanged code on either side of the diff. Bugs hide at the seams.
• For a plan or design doc: trace the proposed flow against the existing system. Where does it touch reality? What does it assume that isn't true?
• Note every place the trace surprises you (unexpected branch, dead code reached, state you didn't know existed). Surprises are signal.

3. Verify — does it actually do what it claims?

For each claim the change/plan makes, answer:

• Does the code path you just traced actually produce that behavior? Walk it explicitly. "It claims X. Path: A → B → C. At C, [observation]. Therefore [holds / doesn't hold]."
• What inputs / states would break it? Edge cases, concurrent callers, error paths, partial failures, retries, empty/null/unicode/huge inputs, ordering assumptions.
• What does it silently change? Performance, error semantics, observability, contract for other callers, on-disk / on-wire format.
• How is it tested? Do the tests actually exercise the traced path, or do they pass while skipping it (mocks that hide the bug, asserts on intermediate state, happy path only)?

4. Report

Output one tight section per finding. Order by severity (blocker → major → nit). For each:

• Finding — one sentence, specific. Cite file:line when applicable.
• Why it matters — the consequence, not the principle.
• Evidence — the trace step or input that exposes it.
• Suggested change — concrete, minimal.

Close with a one-line verdict: ship / fix-then-ship / rework / reject — with the single biggest reason.

Operating rules

• No rubber-stamps. "LGTM" is not an output. If you genuinely find nothing, say what you traced and what you checked, so the user can judge whether your review covered the surface they cared about.
• Cite or it didn't happen. Every claim about the code references a specific path, file, or line. No vague "this might break under load."
• Distinguish claim from verification. "The PR says X" and "I traced X and confirmed / refuted it" are different — keep them separate in the output.
• One simpler-alternative pass is mandatory. Even on small changes, spend one breath asking if the whole thing is necessary. Skip only if the user explicitly says "don't question scope."
• Don't pad with style nits when there's a structural problem. If step 1 or step 2 surfaces a real issue, lead with it; defer nits or drop them.
• No flattery, no hedging. "This is a great PR but..." adds nothing. State the finding.