Menu
The review-first prompt is selected by default. You can switch to a direct command or download a local copy.
Read SKILL.md and any companion files shown by SkillsMP before deciding whether to install.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
npx skills add https://github.com/majiayu000/claude-skill-registry --skill pr-retroThe command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
LLM token logprobs and calibration. Per-decision confidence, ECE, Brier, reliability diagrams, low-confidence triage.
Analyze LLM token logprobs and calibration. Use for per-decision confidence, ECE, Brier scores, reliability diagrams, and low-confidence triage.
回顾最近 N 天的 Claude Code 使用记录——扫描原始会话数据，按主题分组汇总"我都做了什么"，并从个人操作系统视角输出模式、风险与增删建议。当用户说 /recap、"看看我这几天做了什么"、"回顾一下我最近的会话"、"这两天我用 claude 干了啥"、"活动回顾" 时使用。
Based on SOC occupation classification
Showing SKILL.md
| name | pr-retro |
| description | Learn from PR outcomes. |
| practices | ["dora-metrics","sre","lean-startup"] |
| hexagonal_role | supporting |
| consumes | [] |
| produces | [".agents/research/*.md"] |
| context_rel | [] |
| skill_api_version | 1 |
| context | {"window":"fork","intent":{"mode":"task"},"sections":{"exclude":["HISTORY"]},"intel_scope":"topic"} |
| license | MIT |
| compatibility | Requires git, gh CLI |
| metadata | {"author":"AI Platform Team","version":"1.0.0","tier":"contribute","internal":false} |
| allowed-tools | Read, Write, Bash, Grep, Glob |
| output_contract | .agents/learnings/YYYY-MM-DD-pr-*.md |
Learn from PR outcomes by analyzing accept/reject patterns.
After a PR is merged or rejected, analyze what worked and what didn't to improve future contributions.
Output: .agents/learnings/YYYY-MM-DD-pr-{repo}-{outcome}.md
When to Use:
1. PR Discovery -> Find the PR to analyze
2. Outcome Analysis -> Merged/rejected/changes requested
3. Feedback Extraction -> What did reviewers say?
4. Pattern Identification -> What worked/didn't
5. Lesson Extraction -> Reusable learnings
6. Output -> Write retro document
# If PR number provided
gh pr view <number> --json state,reviews,comments,mergedAt,closedAt
# Find recent PRs by you
gh pr list --state all --author @me --limit 10
# Find PRs to a specific repo
gh pr list -R <owner/repo> --state all --author @me --limit 10
| Outcome | Meaning | Focus |
|---|---|---|
| Merged | Success | What worked? |
| Closed (not merged) | Rejected | Why? |
| Open (stale) | Ignored/abandoned | What went wrong? |
| Changes requested | Needs work | What feedback? |
# Get PR outcome
gh pr view <number> --json state,mergedAt,closedAt,reviews
# Get all review comments
gh pr view <number> --json reviews --jq
gh api repos/<owner>/<repo>/pulls/<number>/comments --jq
gh view <number> --json reviews --jq
| Category | Examples |
|---|---|
| Style | Naming, formatting, conventions |
| Technical | Algorithm, architecture, patterns |
| Scope | Too big, scope creep, unrelated changes |
| Testing | Missing tests, coverage, edge cases |
| Documentation | Missing docs, unclear comments |
| Process | Wrong branch, missing sign-off |
| What Worked | Evidence |
|---|---|
| Small, focused PR | < 5 files |
| Followed conventions | No style comments |
| Good tests | No "add tests" requests |
| Clear description | Quick approval |
| What Failed | Evidence |
|---|---|
| Too large | "Please split this PR" |
| Scope creep | "This is out of scope" |
| Missing tests | "Please add tests" |
| Wrong approach | "Consider using X instead" |
## Lesson: [Title]
**Context**: [When does this apply?]
**Learning**: [What did we learn?]
**Action**: [What to do differently?]
**Evidence**:
- PR #N: [quote or summary]
| Lesson | Action |
|---|---|
| PR too large | Split PRs under 200 lines |
| Missing context | Add "## Context" section |
| Style mismatch | Run linter before PR |
| Missing tests | Add tests for new code |
| Slow review | Ping after 1 week |
Write to .agents/learnings/YYYY-MM-DD-pr-{repo}-{outcome}.md
# PR Retro: {repo} #{number}
**Date**: YYYY-MM-DD
**PR**: {url}
**Outcome**: Merged / Rejected / Stale
## Summary
{What was the PR about? What happened?}
## Timeline
| Date | Event |
|------|-------|
| {date} | PR opened |
| {date} | First review |
| {date} | {outcome} |
## Feedback Analysis
### Positive Feedback
- {quote}
### Requested Changes
- {quote}
### Rejection Reasons (if applicable)
- {quote}
## Lessons Learned
### Lesson 1: {title}
**Context**: {when this applies}
**Learning**: {what we learned}
**Action**: {what to do differently}
## Updates to Process
{Any changes to make to pr-prep, pr-plan, or other skills}
## Next Steps
{Future actions based on this retro}
| DON'T | DO INSTEAD |
|---|---|
| Skip retros on merged PRs | Learn from success too |
| Blame maintainers | Focus on what YOU can change |
| Generic lessons | Specific, actionable learnings |
| Skip rejected PRs | Most valuable learning source |
User says: "Run a retro on why this PR was rejected."
What happens:
User says: "Extract what worked from this merged PR."
What happens:
| Problem | Cause | Solution |
|---|---|---|
| Retro is generic | Feedback not tied to evidence | Cite specific comments/decisions and outcomes |
| No clear lesson extracted | Analysis stayed descriptive | Convert observations into behavior changes |
| Maintainer signal is mixed | Contradictory review comments | Separate hard blockers from preference feedback |
| Process changes not adopted | Lessons not operationalized | Add explicit updates to prep/plan/validate workflow |