diff-review

Reviews git diffs, staged changes, and PRs read-only with evidence-based findings.

1|Updated Mar 18, 2026
One-click install
npx skills add https://github.com/psw7205/skills --skill diff-review-psw7205
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: diff-review
Source: https://github.com/psw7205/skills/tree/main/skills/diff-review
Command: npx skills add https://github.com/psw7205/skills --skill diff-review-psw7205

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Code review requests often trigger either shallow skim-reading or expensive multi-agent fan-out, producing false positives or wasted tokens. This Skill reviews a change set (diff, staged changes, recent commits, or a PR) read-only and produces a prioritized findings list where every claim is anchored to a file:line and verified before reporting. ## Core Features & Use Cases - Cost-proportional review ladder: Defaults to L1 (inline reading plus runtime verification of risky claims via repro scripts, type checks, or tests), auto-escalates only to L2 for high-risk diffs, and runs L3 multi-axis fan-out solely on explicit user request. - False-positive filtering: Every finding passes a pre-report gate, quote-the-line evidence requirement, and falsification attempt; severity and confidence are scored separately, and low-confidence findings are withheld. - Structured output and delegation: Findings carry stable IDs, severity, confidence, effort, and autofix labels, and fix requests are handed off to a review-fix-verify loop rather than edited in place. - Use Case: Before merging a pull request, ask for a review and receive a summary header with diff scope and level, plus findings like "#2 [High · conf:high] path/to/file.ext:42" with quoted evidence, impact, and suggested fix. ## Quick Start Review the staged changes in my current git repository and list any high-severity findings with file and line evidence.

Frequently Asked Questions about diff-review

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I review a git diff or pull request with AI?▼

Point the reviewer at a change set using git diff, git diff --staged, git diff <base>...HEAD, or gh pr diff <n>. It reads the changed lines plus their blast radius, then returns severity-ranked findings with file:line evidence and confidence scores.

What is the difference between reviewing a diff and reviewing a whole codebase?▼

Diff review targets only what changed and its blast radius: callers, invariants, and data flow affected by the change. Whole-codebase review scans the entire project across multiple axes like structure, performance, and security, which is a separate, broader workflow.

How does the review avoid false positives?▼

Each finding must pass a pre-report gate requiring a quotable file:line, a named failure mode, and a defensible severity. High and Critical findings need verbatim line quotes, runtime claims are verified with repro scripts or type checks, and low-confidence findings are withheld.

When should I use multi-agent fan-out for code review?▼

Fan-out (L3) is reserved for explicit requests like release-gate audits or security deep dives, since it is expensive and yields diminishing returns. Routine commits and PRs are covered by the default L1 pass, escalating at most to a single isolated reviewer for risky diffs.

Can the review also fix the issues it finds?▼

No, the review is strictly read-only. Fix requests are delegated to a review-fix-verify-commit loop, receiving the selected findings by stable ID, severity, file:line, and improvement direction as the fix scope.