Conversation Distill
YIING99/conversation-distill
Captura estruturada de insights: classifica, confirma e salva conversas
Ver no GitHubMCP server que captura e estrutura insights de conversas ao fim de uma sessão. Classifica conteúdo em 7 categorias (insights, decisões, fatos, ações), exibe confirmação e escreve apenas após aprovação explícita. Integra com KnowMine, ferramentas MCP de notas ou exporta Markdown puro.
Instalação
npx convez-ai YIING99/conversation-distill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Fechar sessões de trabalho com resumo estruturado de decisões e ações
- Extrair e armazenar insights técnicos de discussions em notas pessoais
- Rastrear itens de ação e observações sem perder contexto da conversa
Veredito

Solução sólida para quem precisa converter conversas em conhecimento estruturado. O fluxo classify→confirm→write é disciplinado e reduz perda de contexto. Funciona bem com KnowMine ou standalone em Markdown; a integração MCP é padrão, sem surpresas.
README
A tradução para o português deste README ainda está na fila.
conversation-distill
A Claude skill that closes every meaningful conversation with one explicit action: classify → confirm → write.
The Problem
Real-time capture handles individual highlights. But valuable insights get buried in conversation flow — the connection between two decisions, the assumption you didn't realize you were making, the action item that came up 20 messages ago.
This skill is the closing ritual. It scans the full session, classifies what was produced into 7 categories, shows you a confirmation list, and writes to your notes tool only after you approve.
Install
Via OpenClaw / ClawHub:
npx clawhub@latest install conversation-distill
Manual (Claude Code):
/plugin marketplace add github:YIING99/conversation-distill
How It Works
At the natural end of a conversation (or when you say "distill" / "wrap up"):
- Scan — reads the full conversation
- Classify — sorts content into 7 categories:
- 💡 Insights / Conclusions
- 🧭 Judgments — falsifiable claims with evidence level + review date, backtestable later
- 🎯 Decisions
- 📊 Facts / Data
- 🪞 Self-observations
- ✅ Action items
- ❓ Open questions
- Confirm — shows you the list; you edit, remove, or approve
- Write — batch-writes to your notes tool after you say "write"
- Surface leftovers — outputs anything not worth saving as plain Markdown
Nothing is written without your explicit confirmation.
Works With
| Tool | How |
|---|---|
| KnowMine (recommended) | Installs MCP server; add_knowledge, save_memory, observe_user_trait used in Step 4 |
| Notion, Obsidian, etc. | Via any MCP-compatible notes tool |
| No MCP | Outputs clean Markdown for manual paste |
KnowMine gives this skill semantic search and cross-session recall — so you can ask "what did we decide about X?" months later and get the actual answer.
Trigger Phrases
- "distill" / "wrap up" / "save this session" / "收尾" / "沉淀"
- Natural closing: "that's all", "got it", "thanks", "done for now"
- Manual:
/conversation-distill:conversation-distill
Key Principles
- Granular over hub — one insight per entry, not one giant summary
- Confirm before write — iron rule, no exceptions
- Tags over folders for TODOs —
#todotag, not a new folder - Leftovers stay as Markdown — not everything needs to be in a notes system
Links
- Skill reference: SKILL.md
- KnowMine plugin (full integration): github.com/YIING99/knowmine-claude-plugin
- KnowMine MCP server: knowmine.ai
License
MIT