Review Repository Instructions

Read discovery and retention. Review is read-only by default, including no report file creation. It does not run services or modify host settings merely to verify a documentation claim.

Investigation

Read the complete target and the actual applicable instruction chain. Inspect referenced scripts/configuration and representative consumers before deciding a rule is stale. Verify the working directory and host loading scope, especially when Claude/AGENTS files coexist or nested overrides disagree. If loading cannot be observed, distinguish documented expected loading from observed loading.

Apply the shared seven-dimensional editorial rubric. Use real deletion and collision scenarios, not random line sampling or fixed length thresholds. Distinguish an obsolete fact from an intended constraint the implementation is violating. Retain important rules whose rationale needs clarification rather than silently deleting them. Never add a generic warning or policy just to raise a score, and never use a low score to justify rewriting a whole file.

Finding Shape

For each material finding, state the existing rule and location, applicable host/scope, observed repository evidence, concrete consequence, proposed smallest change, and remaining uncertainty. Separate confirmed drift/conflict, editorial suggestion, and unresolved behavior. Include valuable content to retain when a proposed edit could otherwise lose it. Return one consolidated result, not per-file reports or a grading dashboard. Use optional numeric rubric ratings only when requested, explicitly as editorial judgment, not measured model effectiveness.

Authorized Edits

When the request includes applying improvements, edit only the named scope and preserve useful content/organization. No automatic deletion, import-chain rewrite, policy hierarchy invention, or new instruction-file family. Read back the result and compare against both source evidence and the original meaning. An edit does not prove the host loaded it; record any unverified loading behavior without claiming success. For missing-file creation, use init.

For repository-specific GitHub Copilot Code Review customization, use copilot-review-customizer. That workflow selects Copilot's review surfaces and uses review/regression history; it does not change this skill's Claude/Codex initialization or instruction-review contract.