Distill-Me: Aprender seu estilo em SKILL.md
LewenW/claude-distill-me
Destila seu padrão de decisão em um skill que o Claude pode aprender
Ver no GitHubMCP server que analisa padrões de julgamento, estilo de comunicação e prioridades em sessões do Claude, gerando um SKILL.md que ensina a IA a trabalhar como você. Extrai comportamentos em três camadas: observável, interpretativa e contrastiva, capturando o que realmente diferencia seu modo de trabalho.
Instalação
npx convez-ai LewenW/claude-distill-me1 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
- Criar um perfil de trabalho pessoal reutilizável em todas as sessões
- Ensinar IA a respeitar seu estilo de comunicação e prioridades reais
- Documentar raciocínio e valores implícitos para continuidade de projeto
Veredito

Ferramenta inovadora que vai além de memory tools genéricas, capturando camadas profundas de padrão pessoal. Útil para quem quer consistência em interações com IA, mas exige dados históricos suficientes e conforto com análise de si mesmo para gerar insights reais.
README
A tradução para o português deste README ainda está na fila.
Distill-Me
Distill yourself into a SKILL.md. Learns your judgment, style, and priorities from Claude data.
Skills teach Claude how to do a type of work. Distill-Me teaches Claude how to do work like you.
Most memory tools capture "prefers tabs over spaces." Distill-Me captures "makes decisions by checking evidence before trusting reports, ships fast and iterates, handles ambiguity by demanding honest opinions, and their corrections reveal past bad experiences with over-helpful AI."
How It Works
/distill-me:distill
One command. Scans your Claude sessions, memories, and rules. Extracts behavioral patterns at three depths:
Layer 1: Observable — what you literally say and do
Layer 2: Interpretive — what your patterns reveal about how you think
Layer 3: Contrastive — what makes you DIFFERENT from most people
Output goes to ~/.claude/CLAUDE.md (loads everywhere) and the plugin's skills/enhanced-self/SKILL.md.
Scanned data stays on disk. Analysis is done by Claude (via API), same as any Claude Code session.
Install
git clone https://github.com/LewenW/claude-distill-me.git
cd claude-distill-me
pip install -e .
# Register MCP server
claude mcp add distill-me -e PYTHONPATH="$(pwd)/src" -- python -m distill_me.server
Usage
/distill-me:distill # extract patterns → personal skill
/distill-me:distill pm # same, fused with PM best practices
What Gets Extracted
Three categories, each with three-layer depth:
Decision-Making & Cognitive Patterns
- How you choose when presented with options
- Your implicit decision framework (evidence-driven? gut? speed-first?)
- Risk calibration and when it shifts by context
- What your guardrails reveal about past experiences
- What makes your reasoning different from most people
Communication & Personality
- Voice, language mixing patterns, message length as priority signal
- Your relationship with AI and how it shifts by task
- Emotional register — what triggers excitement, frustration, dismissal
- Gap between how you write and how you want Claude to write
- What makes your voice impossible to confuse with someone else's
Values & Work Philosophy
- Stated preferences vs revealed preferences (where they diverge)
- Values hierarchy — when two priorities conflict, which wins
- Root values driving surface decisions
- Meta-cognition: how you learn, handle your own mistakes
## Decision-Making & Cognitive Patterns
- **Pattern**: When receiving external input (bug reports, reviews), this user
verifies each claim against current code before acting — because they've been
burned by trusting reports that were already handled.
**Evidence**: "很多之前报的 bug 实际上已经在代码中处理了"
**Depth**: interpretive
- **Pattern**: When scoping work, cuts aggressively — only fixes functional +
consistency issues, skips low-impact edge cases. This isn't laziness; it's a
trained instinct that scope creep kills shipping velocity.
**Evidence**: "Bug 1 和 Bug 3 值得修。其余都是极端边缘情况,不动。"
**Depth**: deep
## Communication & Personality
- **Pattern**: Uses 2-5 word commands for routine tasks ("推", "好啦", "帮我弄吧")
but switches to detailed paragraphs when stakes are high. Message length IS
the priority signal — short means "just do it", long means "think carefully."
**Depth**: interpretive
- **Pattern**: Hates AI smell in output. Not just banned words ("leverage",
"robust") — the underlying objection is to THINKING that produces generic,
hedge-everything, please-everyone output. Wants opinionated, specific, direct.
**Depth**: deep
Role Fusion
Fuse your personal patterns with role-specific best practices:
/distill-me:distill pm # built-in PM template
/distill-me:distill discord/bot-builder # from installed Cowork plugin
Two sources of roles:
- Built-in templates —
references/role-templates/*.md(ships with PM) - Installed plugins — any Cowork plugin in
~/.claude/plugins/with a SKILL.md
Install a Cowork plugin, and its skills automatically become available as roles. Personal patterns always take priority — the role template adapts to fit you, not the reverse.
Configuration
All limits are configurable via environment variables:
| Variable | Default | What it controls |
|---|---|---|
DISTILLME_SCAN_DAYS | 30 | How many days of history to scan |
DISTILLME_MAX_SESSIONS | 20 | Max session files to read |
DISTILLME_MAX_TOTAL_TURNS | 100 | Max turns to collect |
DISTILLME_MAX_ANALYSIS_TURNS | 80 | Turns sent to analysis |
DISTILLME_MAX_MESSAGE_CHARS | 2000 | Per-message truncation |
DISTILLME_MAX_MEMORY_CHARS | 1000 | Per-memory truncation |
DISTILLME_EXCLUDE_PROJECTS | (none) | Comma-separated project names to skip |
Set them in your .mcp.json env block or shell profile.
Data Sources
| Source | Location |
|---|---|
| Session logs | ~/.claude/projects/*/*.jsonl |
| Memory files | ~/.claude/projects/*/memory/*.md |
| CLAUDE.md | ~/.claude/CLAUDE.md + project-level |
| memory-bridge | ~/.claude/shared-memory/ |
MCP Tools
| Tool | Purpose |
|---|---|
scan_user_data() | Collect all data, return three-layer analysis prompts |
save_extracted_patterns() | Write judgment/style/priorities to disk |
generate_personal_skill() | Build SKILL.md from patterns + optional role fusion |
Structure
├── commands/distill.md # /distill-me:distill command
├── skills/
│ ├── self-distill/SKILL.md # extraction quality guide
│ └── enhanced-self/SKILL.md # [generated] your personal skill
├── references/role-templates/pm.md # PM best practices
└── src/distill_me/
├── server.py # MCP server (3 tools)
├── scanner.py # data collection
├── extractor.py # three-layer extraction prompts
└── generator.py # SKILL.md + CLAUDE.md injection
Relation to memory-bridge
memory-bridge stores and shares memories across projects. Distill-Me reads that data and goes further — it extracts behavioral patterns.
memory-bridge remembers what you said. Distill-Me learns how you think.
Platform Compatibility
Works everywhere Claude loads ~/.claude/CLAUDE.md:
| Feature | Claude Code CLI | Desktop | Cowork |
|---|---|---|---|
| MCP tools | yes | yes | needs bridge |
/distill-me:distill | yes | yes | yes |
| Output to CLAUDE.md | yes | yes | yes |
Requirements
- Python 3.10+
mcp>=1.0.0- Claude Code (CLI or Desktop)
License
MIT