| name | scrutinizing-code-changes |
| description | Outsider-perspective end-to-end review of a plan, PR, or code change. First questions intent and whether a simpler/more elegant approach would achieve the same goal, then traces the actual code path (not just the diff) to verify correctness, style conventions, and best practices. Output is concise, actionable, and constructive. Trigger on /scrutinize and proactively whenever the user asks to review, audit, sanity-check, or get a second opinion on a plan, PR, diff, design doc, or proposed code change. |
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.
- Constructive & Polite. Balance rigorous technical scrutiny with professional courtesy, highlighting genuine strengths and positive aspects.
- 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 & Summary — what is this actually trying to do?
- Provide a concise summary of the proposed changes.
- 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)?
- What standards or style conventions are violated? Check for language-specific standards (e.g., PEP 8 for Python, Google Style Guide).
- What best practices could be improved? Assess readability, maintainability, modularity, and opportunities to simplify redundant structures.
4. Report
Structure your report clearly as follows:
- Executive Summary: A concise summary of the changes and overall assessment.
- Strengths & Positive Aspects: Highlight genuinely elegant solutions, robust test coverage, or clean patterns.
- Key Issues & Improvements: One tight section per finding, ordered 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."
- Substance over Minor Nits. Focus on substantial issues, logic, correctness, and best practices. Do not pad with minor style nits unless they violate a strict project style guide.
- Constructive Politeness over Empty Flattery. Highlight genuine, tangible positive aspects of the code as strengths, but keep feedback technically direct and professional.
- Escalate Ambiguity. If the code is too complex, ambiguous, or lacks context, stop and ask the author for clarification.
