# Change Review > Review a diff, branch, PR, or uncommitted changes for correctness and maintainability problems. Exclude whole-codebase audits and implementation without a review request. - Skill: `stevennitesh/change-review` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add stevennitesh/change-review` - Raw SKILL.md: https://api.skillmd.com/api/skills/stevennitesh/change-review/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: stevennitesh (https://skillmd.com/u/stevennitesh) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/stevennitesh/change-review --- # Change review Judge whether a change delivers the intended behavior and whether its implementation is sound. Review without editing product code or publishing external comments. For a review-and-fix request, finish the review phase before returning findings to the authorized repair work. Review conclusions do not grant merge or release authority. Use ordinary review by default. Only when the user explicitly requests high assurance or a review by multiple independent reviewers, read [High assurance](references/high-assurance.md). A large diff, a PR, or the phrase "final review" alone does not activate that mode. Ordinary review needs no fanout. ## 1. Identify what is being reviewed Use the target and comparison the user names. Otherwise use relevant current WIP, including staged, unstaged, and in-scope untracked content. For a named branch or PR, resolve its actual base and head; do not assume a branch called main or silently replace an exact comparison with a merge-base comparison. If the target cannot be inferred, ask for the missing selection. An empty selection means no changes to review, not a successful review of an imagined candidate. Record the resolved comparison and candidate identity. For mutable work, capture the selected content, including untracked files and relevant surrounding evidence, or hold exclusive custody while reviewing. A diff alone is insufficient when callers, configuration, or stored representations determine behavior. Bind the context and supplied proof to the version actually examined. Read the request and governing requirements, repository guidance, and relevant accepted decisions. Use a spec when one governs the work; do not require a tracker or fabricate intent from tests or implementation. Missing intent may limit conformance review while independently evidenced correctness checks can continue. State that limit rather than implying all accepted behavior was verified. When independence matters, identify whether the reviewer authored the change or shares its implementation context. Self-review can be useful but does not satisfy an independent-review requirement; report an unmet requirement to the caller. ## 2. Trace behavior and engineering quality Check the outcome and scope first: does the ordinary caller receive the requested meaning, including relevant rejection, partial-success, and completion behavior? Then independently assess correctness, ownership, representation, simplicity, maintainability, and proof. Working happy-path behavior does not excuse a concrete design cost; attractive structure does not excuse an incomplete outcome. If the implementation follows a plan whose technical assumptions fail the accepted outcome, identify the plan-level problem and affected decision. Do not demand faithful implementation of a flawed mechanism or silently redefine the user's goal. Trace meaningful changes through real callers, owners, and effects. When a result crosses stages or tickets, check that its actual produced or persisted form reaches the consumer and preserves accepted values, issues, identity, and terminal behavior. Follow serialization, configuration, dependency versions, lifecycle, and cross-language consumers when they matter even if absent from the visible call graph. For numerical or data transformations, check material units, time alignment, missing-value meaning, precision, and aggregation assumptions. Where correctness depends on the method, seek an independent reference, analytic case, or invariant rather than reproducing the implementation. Inspect activated risks: trust and authorization boundaries, shared state, resource bounds, recovery, migration, and independently deployed consumers. For removed or replaced behavior, check affected registrations, configuration, callers, tests, and public guidance. Preserve necessary compatibility rather than demanding deletion merely because two paths temporarily coexist. Challenge duplicated decisions, exposed internals, unnecessary state, and layers that add caller burden without useful ownership or policy. Check whether a simpler repository-native approach preserves the required meaning. File size, one adapter, unfamiliar syntax, or repeated text alone does not establish a design defect. Keep unrelated baseline improvement discovery with audit-codebase. ## 3. Test finding hypotheses Before reporting, apply [Finding standards](references/finding-standards.md). Trace a concrete affected scenario and seek disconfirming evidence. Use the baseline to distinguish an introduced or worsened problem from unrelated old code. An unchanged line can still be the causal location of a regression activated by the change; explain the connection rather than restricting investigation to hunks. Reuse valid candidate-bound proof. Run safe, proportionate checks when they settle a material question or satisfy repository policy. Scratch harnesses and caches are acceptable within an isolated review; do not mutate live systems or the reviewed source. Check whether tests distinguish a plausible wrong result and whether a substitute bypasses the property under review. Test counts and reviewer agreement are not evidence of correctness. ## 4. Return an evidence-backed conclusion Recheck mutable candidate and decisive context identity. If it moved, identify what was reviewed and what remains unreviewed; do not attach a clean verdict to the new state. A later review can reuse unaffected evidence after revalidation. On a repair review, track prior findings as resolved, still present, disproved, or unresolved. Inspect the repair and its affected consumers for new regressions; broaden to the full successor when the change grows beyond the bounded repair. Return the reviewed identity, decisive checks, material coverage or evidence limits, and actionable findings in impact order. Each finding names precise locations, scenario, evidence, consequence, and the required correction or proof. When no findings remain, say so without implying complete coverage if required evidence is absent. For an explicit gate decision, use Finding standards; distinguish required corrections, nonblocking findings, and any residual-risk acceptance still needed from its owner. Return control to the caller; publication, repair, or delivery follows existing authorization rather than being started by the review itself.