Whetstone

Distill a finished dev session into proposals for a self-growing local skill library. Mines the transcript + git diff + pitfalls, decomposes what was learned into 4 layers (L1 principle / L2 method+pitfall / L3 platform-params / L4 state) via references/extraction-framework.md, reconciles against existing skills (dedup+write-back / supersede / conflict), and writes a human-reviewed proposal into inbox/. Self-contained, zero runtime dependency. Trigger words: 蒸馏这次 / distill this session / 把这次经验做成 skill / 开发完记录经验 / capture what I learned. Use after finishing a feature or bugfix to capture the re

1 starsWriting & Content

Install

skills CLI (Claude Code, Cursor, Codex and more)

npx skills add TbusOS/whetstone --skill whetstone

Manual install

git clone --depth 1 https://github.com/TbusOS/whetstone.git /tmp/whetstone
mkdir -p ~/.claude/skills
cp -r /tmp/whetstone/. ~/.claude/skills/whetstone

Personal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.

From the SKILL.md

# Whetstone · 磨刀石 — Distiller 提炼流程 (Session → Skill)

> 配套 `references/extraction-framework.md`(L1-L4 分层 schema)。本文件管「怎么从一坨会话里把 schema 套出来」。
> **Runtime 中立 + 零运行时依赖**:本文件是 Whetstone 蒸馏器的 Agent Skill 入口,任何 skills-compatible runtime(Claude Code / Codex / Cursor …)都能调。产物是可移植 skill 包(见 `spec/skill-package.md`)。engram(同步)/ llm-wiki(发布)/ darwin-skill(打分)均为**可选** sink,装不装都能跑。

把一次**已经发生**的开发会话(transcript + git diff + 踩坑),
按 extraction-framework 拆成 L1-L4,对账已有 skill 库,产出**提案**交人审,
人点头后才进库。

**和 nuwa 的根本区别:nuwa 去网上调研一个人;本流程是「会话事后挖掘」——素材已经在那,挖出信号、分层、对账、入库。**

---

## 核心原则(贯穿全程)

1. **挖掘,不创造** —— 只提炼会话里真实发生的工程学习,不脑补、不把 AI 自己的元对话当知识。
2. **diff 说 WHAT,transcript 说 WHY + 什么失败了** —— 最终 diff 抹掉了试错过程,坑和决策只在 transcript 里,两个都要挖。
3. **对账库存是命门** —— 大多数会话对应的 skill 已经存在,提炼的真正难点不是"写新的",是"和已有的合并/去重/替代",这也是库漂移的发生地。
4. **提案,不直写** —— 产物是对库的 diff,落 inbox 暂存;**人审 gate 后**才进正式库(SkillLens(Microsoft,https://microsoft.github.io/SkillLens/):无引导的 LLM 评委在两个 skill 里挑更好的那个,准确率 46.4%,与掷硬币无异,所以不许自动覆盖)。

---

## Phase 0 · 圈定范围

```
1. 确认要提炼哪次任务(刚做完的 feature)
2. 锁定素材:
   - 会话 transcript(哪段)
   - git 范围(commit 区间 / diff)
   - 踩坑 journal 条目(若 capture hook 攒了)
3. 判定领域:这次是关于什么?(安全启动 / 某驱动 bringup / SDK 迁移 …)
   → Phase 3 对账要用领域定位到对应 skill
4. 信号量筛(阈值):
   - 改了实质东西 / 踩了非平凡的坑 / 做了设计决策 → 继续
   - 纯 typo、纯机械改、零学习 → 不提炼,只往 journal 追加一行,结束
```

**🔴 失败分支:** 一次会话跨多个领域(同时碰了安全启动 + 一个驱动)→ Phase 1 起按领域分流,各自独立走完后续,**不混在一个提案里**。

---

## Phase 1 · 多角度挖掘(对同一会话做多镜头扫描)

不是 nuwa 的并行网搜,是对**同一份会话**用 4 个镜头各扫一遍,捞"学习时刻"。

Read the full skill on GitHub

Details

Skill name: whetstone
Repository: TbusOS/whetstone
License: MIT
Repo updated: 20 October 2026
Listed in: karanb192/awesome-claude-skills

FAQ

What does the Whetstone skill do?

How do I install the Whetstone skill?

skills CLI (Claude Code, Cursor, Codex and more): npx skills add TbusOS/whetstone --skill whetstone. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.

Is Whetstone popular?

It lives in TbusOS/whetstone, which has 1 GitHub stars and 0 forks, and is included in 1 curated skill list.

Related skills

Hitpost

Lihaila/hitpost-skill

Write X (Twitter) posts that earn attention and sound human, using the writing rules from HitPost (lihaila.si). The skill turns an idea, topic, link or rough note into three ready-to-post drafts with different angles (single post, thread or long Premium post, in any tone or language), checks their length the way X counts it, and gives one-click share links. It can also polish and score an existing draft. Use it whenever the user wants to write, draft, rewrite, improve or punch up a tweet, X post, thread or 🧵, or a gm/greeting post, or asks "what should I tweet about…", even if they don't say

Writing & ContentIn 1 list

Publora Post Ideas

publora-team/publora-post-ideas

Suggests three different angles for a social post and writes the one you pick, instead of guessing at a single draft. Use when the request is open-ended — "write me a post", "I need content ideas", "I don't know what to post today", "give me something for LinkedIn" — or when a draft was rejected and a different approach is needed. Pairs with publora-publishing, which does the actual scheduling. Works on its own; with Publora connected it also picks angles against the user's real feed and schedules the result.

Writing & ContentIn 1 list

Mindpalace

aashutosh396/mindpalace-skill

Resource-first knowledge vault for an AI agent. Single source of truth per resource: infra/ for servers, hosting and DNS; accounts/ for platform logins per owner or client; projects/ as thin pointer files that link to infra and accounts via YAML frontmatter relations. Plus runbooks/ for repeatable ops, docs/ for artifacts, and LOG.md as the raw timeline. Use this skill when the user asks about credentials, secrets, API keys, server URLs, ports, deploy steps, account logins, past decisions or project docs; when scaffolding a new project, server or account entry; or when recording knowledge that

Writing & ContentIn 1 list

Goethecoach

janosszaboaipm-design/goethecoach-claude-skill

German exam coach for the Goethe-Zertifikat (A1–C2) by GoetheCoach. Generates exam-style Lesen, Schreiben and Sprechen practice tasks, evaluates learner answers against the official criteria (Aufgabenerfüllung, Kohärenz, Wortschatz, Strukturen), explains errors through the learner's native language, runs follow-up tutoring and builds study plans up to the exam date. Answers questions about the exam (format, scoring, Leitpunkte, dates, digital exam, Goethe vs. telc, failing) and everyday German for life in Germany from the bundled GoetheCoach article library, with links for further reading. Use

Writing & ContentIn 1 list

Say It Plainly

adjustleads/provenskills-free-packs

Use when a draft sounds like an AI wrote it, or someone asks to make it less AI, more human, plain English, or like them. Rewrites it, facts unchanged.

Writing & ContentIn 1 list

Apify Linkedin Post Engagement

johnisanerd/claude-skill-linkedin-post-engagement

Analyze LinkedIn post engagement with the Apify LinkedIn Posts API Actor (johnvc/linkedin-posts-api). Collect a profile's posts or a list of post URLs, then measure reactions, comments, and shares alongside author reach (followers, headline) to see what performs. Use when the user wants to analyze LinkedIn post engagement, measure engagement rate, find a creator's or a competitor's top-performing posts, benchmark engagement across profiles, track post engagement over time, or build a LinkedIn engagement report. Returns one row per post with numLikes, numComments, numShares, datePosted, and a r

Writing & ContentIn 1 list

Browse the most popular Claude skills, all writing & content skills, or write your own with the skills generator.