Retrospect
Session retrospective that scans the current conversation for agent failures, user corrections, and alignment gaps, then proposes patches to CLAUDE.md or skills. Use when: after a difficult session, after repeated corrections, to improve agent behavior, to analyze what went wrong, to close the learning loop after autopilot/fix-pr. Trigger phrases: "retrospect", "what went wrong", "session retro", "learn from this", "what should we improve", "session review", "capture learnings", "post-mortem this session".
Install
/plugin marketplace add npow/claude-skillsThe repo is a Claude Code plugin marketplace. Run /plugin afterwards to pick the plugin that includes this skill.
npx skills add npow/claude-skills --skill retrospectgit clone --depth 1 https://github.com/npow/claude-skills.git /tmp/claude-skills
mkdir -p ~/.claude/skills
cp -r /tmp/claude-skills/retrospect ~/.claude/skills/retrospectPersonal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.
From the SKILL.md
# Session Retrospect Scan the current conversation for alignment failures, classify root causes, and propose durable instruction patches. This is NOT a team retro — it analyzes agent behavior in THIS session and produces fixes that prevent recurrence across all future sessions. ## Workflow ``` SCAN → CLASSIFY → DIAGNOSE → PRESCRIBE → APPLY ``` ### Phase 1: SCAN — Detect Failure Signals Sweep the conversation for signals at six priority levels. Gather evidence before asking questions. **P1 — Explicit corrections (highest confidence):** User directly identified a problem: - Correction language: "no", "not that", "wrong", "incorrect", "that's not what I meant" - Halt commands: "stop", "don't do that", "undo", "revert" - Repeated instructions (user re-explained something already said) **P2 — Agent self-corrections:** - "let me try again", "I see the issue", "I made a mistake" - Same tool called 3+ times with varying parameters (floundering) - Approach abandoned mid-execution and pivoted **P3 — Structural failures:** - Tool calls that returned errors - Required skill not invoked when it should have been - Verification step omitted (tests, build, diff check) - Premature completion claims without evidence - Unauthorized scope expansion (touched unrelated files) - Silent degradation after tool failure (switched to guessing) **P4 — Implicit misalignment (verify before reporting):** - User tone shift toward more directive language
Details
- Skill name
retrospect- Repository
- npow/claude-skills
- Path
retrospect/SKILL.md- License
- Apache-2.0
- Repo updated
- 20 August 2026
FAQ
What does the Retrospect skill do?
Session retrospective that scans the current conversation for agent failures, user corrections, and alignment gaps, then proposes patches to CLAUDE.md or skills. Use when: after a difficult session, after repeated corrections, to improve agent behavior, to analyze what went wrong, to close the learning loop after autopilot/fix-pr. Trigger phrases: "retrospect", "what went wrong", "session retro", "learn from this", "what should we improve", "session review", "capture learnings", "post-mortem this session".
How do I install the Retrospect skill?
Claude Code plugin: /plugin marketplace add npow/claude-skills. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.
Is Retrospect popular?
It lives in npow/claude-skills, which has 5 GitHub stars and 1 forks.
Related skills
Autopilot
5npow/claude-skills
Use when the user says "autopilot", "build this", "make it", "ship it", "just do it", "do it", "implement this", or gives a vague idea, spec, task list, or bug fix and expects autonomous execution. Handles decomposition, agent coordination, and verification automatically. "build me end to end", "go from idea to code", "do everything", "run this autonomously". Follows the Spec-Driven Development workflow: specify → clarify → plan → tasks → analyze → implement.
Build Python Library
5npow/claude-skills
Implements pip-installable Python libraries from scratch using src layout, pyproject.toml, and pytest. Use when the user asks to build, create, implement, or scaffold a Python library, package, pip package, middleware, SDK, or wrapper. Covers project scaffolding, proxy/wrapper patterns, provider-agnostic interfaces, async/sync support, and test infrastructure.
Canon Design Review
5npow/claude-skills
Architecture-level design review evaluating 15 Tier 2 best practices drawn from DDD (Evans), POEAA (Fowler), Release It! (Nygard), DDIA (Kleppmann), Team Topologies (Skelton/Pais), How Buildings Learn (Brand), and Systemantics (Gall). Auto-invoke when a PR touches: - Service definitions, infrastructure code, API boundaries, or database schemas - Files matching **/architecture*, **/design*, **/api/*, **/schema*, **/migrations/* - Any directory named services/, infrastructure/, platform/, domain/, core/ Also invoke when the user says: - "design review", "architecture review", "review the design"
Seedance Director
5GAGAA-HUE/ASHUO-claude-skills
电影级视频分镜头提示词生成器,专为 Seedance 2.x 系列模型设计。当用户需要生成视频、制作分镜脚本、描述影像画面、进行镜头调度、输出 AI 视频提示词(prompt)时触发。默认输出适合直接复制给模型的轻量成品提示词,优先贴近长镜头、多分镜、蒙太奇三类实用模板。动作戏默认采用事件链长镜头逻辑;变身、异变、强烈招式前摇等场景采用阶段推进逻辑。多分镜默认强制机位四维轮换(景别 / 角度 / 视点 / 运动方式)以避免构图单调,并严格守恒 180 度轴线以避免越轴。仅当用户明确要求工业版、完整版、分镜表或后期说明时,才输出重型结构化版本。
Prompt Clash Ensemble
5npow/claude-skills
Fans out 4 parallel prompt-clash defend agents at staggered time budgets (1min, 2min, 3min, 4min), then ensembles/synthesizes the best elements of each into a single hardened prompt. Use when the user wants the strongest possible defense prompt and has ~5 minutes. Trigger phrases include "prompt clash ensemble", "ensemble defense", "fan out prompt clash", "best defense possible", "ensemble prompt", "multi-budget defense".
Suno Prompter
5GAGAA-HUE/ASHUO-claude-skills
Suno v5.5 Pro 音乐提示词 + 歌词生成专家——将用户提供的任意简单需求(情绪、场景、歌词方向、用途)转化为符合 Suno v5/v5.5 规则的完整方案:Style Prompt(英文风格提示词)+ 专业歌词(含韵脚/节奏/音节控制)+ 负向约束。当用户提到 Suno、AI 音乐生成、帮我写提示词、帮我生成歌曲、我想用 Suno 做一首歌、Suno prompt 怎么写、风格提示词、歌词、作词、帮我写歌、AI 作曲等任何场景时,立即触发此 skill,主动介入并输出完整可粘贴方案。即使用户只是说"我想做一首悲伤的歌"也应直接生成,不要先问用户要不要使用此技能。
Browse the most popular Claude skills, all ai, llms & agents skills, or write your own with the skills generator.