Self-Review · wrg32786 — Claude Code skill: what it does, how to install
The Self-Review Claude Code skill sends substantial code changes through a focused second review. It targets multi-file edits, hot paths, prompts, and work that still feels uncertain.
MIT ★ 26 updated 2026-11-16
npx skills add wrg32786/aigent-os --skill self-review -a claude-codeWhat Self-Review does
Self-Review prepares a concise account of the request, the cause of the issue, and the change made. It lists affected files, then asks focused questions tied to a function or line and a possible failure. Edge cases and interactions with code outside the change are part of the review too.
The skill passes this context to a code-reviewer subagent through the Agent tool and requests HIGH, MEDIUM, or LOW ratings with line references. HIGH findings must be fixed before the work is declared done; decisions on MEDIUM and LOW findings are recorded with their reasons. A mechanical one-line edit can be skipped, but the skip must be stated. Running the review requires the Agent tool and a code-reviewer subagent.
What it helps with
- review a change spanning several files
- look for regressions in a hot path
- check edge cases after a code change
- assess a reviewer’s disputed finding
Example prompts for the agent
- Review this change separately and identify possible regressions with line references.
- Check the edits across these files for edge cases and unintended interactions.
- Before finishing, ask the
code-reviewersubagent to inspect the change.
How to install Self-Review
1Skills CLI (Vercel)
npx skills add wrg32786/aigent-os --skill self-review -a claude-code2Manually: put the files in place
~/.claude/skills/self-review/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/self-review/
1Skills CLI (Vercel)
npx skills add wrg32786/aigent-os --skill self-review -a codex2Manually: put the files in place
.agents/skills/self-review/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add wrg32786/aigent-os --skill self-review -a cursor2Manually: put the files in place
.cursor/skills/self-review/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install wrg32786/aigent-os skills/self-review2Manually: put the files in place
.github/skills/self-review/SKILL.md1Skills CLI (Vercel)
npx skills add wrg32786/aigent-os --skill self-review -a gemini-cli2Platform CLI
gemini skills install https://github.com/wrg32786/aigent-os --path skills/self-reviewManually: put the files in place
.devin/skills/self-review/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add wrg32786/aigent-os --skill self-review -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about Self-Review
How do I install Self-Review in Claude Code?
Run npx skills add wrg32786/aigent-os --skill self-review -a claude-code. The skill file is placed at ~/.claude/skills/self-review/SKILL.md.
What does Self-Review do?
It prepares targeted questions about failure modes, edge cases, and regressions, then sends them to a reviewer subagent. Findings receive severity ratings and code-line references.
What changes should trigger Self-Review?
It is meant for changes over 30 lines, multi-file edits, hot paths, user-facing prompts, or work that leaves genuine uncertainty. A trivial one-line edit can be skipped if the skip is stated explicitly.
Source text
SKILL.md · MIT
---
name: Self-Review
description: "Spawn a reviewer with specific verification questions. Use when reviewing code changes over 30 lines, across files, in hot paths or prompts, or with uncertainty."
---
# Self-Review
Run an explicit second-pair-of-eyes pass on non-trivial work before declaring done. Quality of review depends on quality of question.
## When to RUN self-review
- More than ~30 lines OR multi-file change
- Hot path (many requests flow through this code)
- Subtle interactions with code you might have missed
- Prompt or output affecting user-visible content
- Code not worked in recently (mental model might be stale)
- Any genuine uncertainty after writing the change
## When to SKIP self-review
- One-line edit, mechanical fix, obvious diff
- Pure rename or move with no logic change
- Code just written with full context, low blast radius
- Cost-benefit doesn't justify (review takes longer than the fix did)
If skipping, say so explicitly: "Skipping self-review — N-line trivial fix." Don't silently bypass.
## How to execute
### Step 1 — Set context (3-5 lines max)
State, in plain language:
- The bug / feature / what was asked
- The cause (what was actually wrong)
- The fix (what was actually changed)
### Step 2 — List specific changes
```
- path/to/file_one.ts:L123-L145 — replaced regex with structured parser
- path/to/file_two.ts:L56 — added null guard for empty corpus
- path/to/config.json — bumped diversity_threshold 0.6 → 0.75
```
### Step 3 — Ask 4-6 specific verification questions
**Bad asks (do not produce useful answers):**
- "Is this good?"
- "Does this look right?"
- "Any issues?"
- "Did I miss anything?"
**Good asks (concrete, falsifiable, named):**
- "Does the new keyword order ever cause a wrong branch in `routeRequest`?"
- "Does the diversity rule reach the model in the same prompt as the recent-titles list?"
- "Is the early return at line 47 still bypassing this check?"
- "If `recent_titles` is empty (first run), does the sort comparator return stable output?"
- "Does the schema migration handle the existing 12K `null` rows or only new inserts?"
Each question must:
- Reference a specific file:line or function name
- State the failure mode you're worried about
- Be answerable yes/no after a focused read
### Step 4 — Ask explicitly about edge cases, regressions, missed interactions
```
"Edge cases I'm uncertain about:
- What happens when input X is empty?
- Does this break callers Y and Z that I didn't read?
- Any regression on the prior fix from PR #N?"
```
### Step 5 — Request severity ratings + line references
```
"Please rate findings as HIGH / MEDIUM / LOW with concrete line references."
```
### Step 6 — Spawn the reviewer subagent
Use the Agent tool with a code-reviewer subagent_type (or equivalent). Pass the full context block (steps 1-5) as the prompt. Set model to whatever is appropriate for the review depth.
### Step 7 — Process findings
- **HIGH severity** → fix before declaring done
- **MEDIUM severity** → fix unless you have a specific reason not to. If skipping, document why.
- **LOW severity** → judgment call. Often fold into a follow-up note.
- **Reviewer misread** → trust your own reading, but re-verify before dismissing. A misread is a signal the code might need clarity.
The reviewer is a **second pair of eyes, not an authority.** Evaluate every suggestion. Blindly applying every note is as bad as ignoring them all.
## Output format
After review completes, append to your delivery message:
```
**Self-review:** {APPROVED — severity X} / {N findings applied} / {N deferred — listed below}
- {finding} → {fix applied OR reason deferred}
```
© wrg32786 · MIT · wrg32786/aigent-os
Similar skills
All skills for Code review and refactoring →Self-Review by other authors comes first.