一键在 Manus 中运行任何 Skill

retrospective

星标215

分支26

更新时间2026年4月29日 20:15

Analyze past Claude Code session(s) for mistakes and propose targeted improvements. Use when the user asks for a retrospective, post-mortem, or session review. Outputs concise findings plus suggested edits to CLAUDE.md (general lessons) or new/updated skills (specific workflows).

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

ItsBenCodes

ItsBenCodes/cookies

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

Retrospective

Review session transcripts, list mistakes, propose improvements.

1. Resolve scope

Parse $ARGUMENTS:

empty → current session — analyze from your existing conversation context. Do not read any jsonl file; the conversation is already loaded.
contains a number + d/day(s) → past sessions modified within that many days. Read jsonl files from ~/.claude/projects/<encoded-cwd>/:
```
SESSION_DIR=~/.claude/projects/$(pwd | sed 's|/|-|g')
find "$SESSION_DIR" -name '*.jsonl' -mtime -N
```
remaining text → extra focus/context to bias the analysis (e.g. "focus on test failures")

2. Analyze

Current session: scan your own conversation context. Extract user corrections, retried tool calls, reverts, "no"/"don't"/"stop", failed builds/tests, repeated identical commands, dead-end paths.

Past sessions (N days): spawn one Explore subagent per jsonl in parallel (single message, multiple Agent calls). Brief each:

Read <path>. List only the mistakes Claude made: user corrections, wrong assumptions, repeated failures, wasted tool calls, friction points. One bullet each, ≤15 words. No fixes, no praise. Under 200 words.

Then reconcile: dedupe, group by theme, keep frequency counts.

3. Report

Output to the user in this shape (no preamble):

## Mistakes
- <mistake> (×N)
- ...

## Improvements
**CLAUDE.md** (general, always-on):
- <rule> — <one-line why>

**Skill: <name>** (specific workflow):
- Trigger: <when to invoke>
- Body: <2-3 bullet behavior>

## Apply?

Routing rule:

CLAUDE.md = always-relevant rules (style, conventions, repo invariants).
Skill = workflow that fires on a specific trigger (slash command, task type).

Wait for confirmation before editing CLAUDE.md or creating skill files.

4. Apply (on confirmation)

CLAUDE.md edits: append/modify in place, keep terse.
New skill: write .claude/skills/<name>/SKILL.md with frontmatter (name, description, optional argument-hint).

Stop. Do not auto-commit.

同仓库更多 Skills

同仓库

publish-release

ItsBenCodes/cookies

Create a GitHub release and publish to NPM.

2026-04-29215

commit

ItsBenCodes/cookies

Commit staged work, push the branch, and open a PR. Use when wrapping up changes that need to ship.

2026-04-29215

create-plan

ItsBenCodes/cookies

Draft an implementation plan for a feature in .agents/plans. Does not execute the plan. Use when the user asks to plan, design, or scope a feature without building it.

2026-04-29215

execute-plan

ItsBenCodes/cookies

Execute an implementation plan. Implements every task, runs validation, commits, and opens a PR. Use when the user asks to execute, implement, or run a plan file.

2026-04-29215

fix-vulnerabilities

ItsBenCodes/cookies

Triage and fix dependency vulnerabilities in an npm project. Use when the user asks to fix, address, or patch dependency/dependabot/npm audit vulnerabilities.

2026-04-29215

prepare-release

ItsBenCodes/cookies

Prepare a release PR by bumping package versions (patch, minor, or major).

2026-04-29215

name	retrospective
description	Analyze past Claude Code session(s) for mistakes and propose targeted improvements. Use when the user asks for a retrospective, post-mortem, or session review. Outputs concise findings plus suggested edits to CLAUDE.md (general lessons) or new/updated skills (specific workflows).
argument-hint	[N days] [extra context]

Retrospective

Review session transcripts, list mistakes, propose improvements.

1. Resolve scope

Parse $ARGUMENTS:

empty → current session — analyze from your existing conversation context. Do not read any jsonl file; the conversation is already loaded.
contains a number + d/day(s) → past sessions modified within that many days. Read jsonl files from ~/.claude/projects/<encoded-cwd>/:
```
SESSION_DIR=~/.claude/projects/$(pwd | sed 's|/|-|g')
find "$SESSION_DIR" -name '*.jsonl' -mtime -N
```
remaining text → extra focus/context to bias the analysis (e.g. "focus on test failures")

2. Analyze

Current session: scan your own conversation context. Extract user corrections, retried tool calls, reverts, "no"/"don't"/"stop", failed builds/tests, repeated identical commands, dead-end paths.

Past sessions (N days): spawn one Explore subagent per jsonl in parallel (single message, multiple Agent calls). Brief each:

Read <path>. List only the mistakes Claude made: user corrections, wrong assumptions, repeated failures, wasted tool calls, friction points. One bullet each, ≤15 words. No fixes, no praise. Under 200 words.

Then reconcile: dedupe, group by theme, keep frequency counts.

3. Report

Output to the user in this shape (no preamble):

## Mistakes
- <mistake> (×N)
- ...

## Improvements
**CLAUDE.md** (general, always-on):
- <rule> — <one-line why>

**Skill: <name>** (specific workflow):
- Trigger: <when to invoke>
- Body: <2-3 bullet behavior>

## Apply?

Routing rule:

CLAUDE.md = always-relevant rules (style, conventions, repo invariants).
Skill = workflow that fires on a specific trigger (slash command, task type).

Wait for confirmation before editing CLAUDE.md or creating skill files.

4. Apply (on confirmation)

CLAUDE.md edits: append/modify in place, keep terse.
New skill: write .claude/skills/<name>/SKILL.md with frontmatter (name, description, optional argument-hint).

Stop. Do not auto-commit.