Mine old chats into reusable
skills you own permanently
You've already trained Claude your preferences across dozens of conversations. One extraction prompt recovers all of it — preferences, rejections, workflow patterns, style signals — and packages them into a permanent skill.md you can drop into any future project.
Recommended Models
This skill needs a long-context model to read entire conversations. Don't use flash/mini models — they truncate context and miss early signals.
| Model | Best For | Cost | Quality | |
|---|---|---|---|---|
| Claude Sonnet 4.6 | Deepest analysis — catches nuance and implicit rules | Moderate | ★★★★★ | Recommended |
| Claude Haiku 4.5 | Fast extraction on shorter chats at low cost | Cheap | ★★★★★ | Recommended |
| GPT-5.4 | Good alternative for OpenAI-centric workflows | Moderate | ★★★★★ | Recommended |
Temperature tip: Set to 0.3. This is extraction work, not creative generation — you want consistent, faithful analysis of what's actually in the conversation.
Cost Estimate
💚 Cheap Tier
~$0.01
per extraction, on Claude Sonnet 4.6
~2,000
input tokens / run
~800
output tokens / run
One-time
per conversation
One-time cost per conversation. The skill.md you get is free to reuse forever. Longer conversations cost slightly more — a 6-month project chat might run 8,000+ input tokens, still well under $0.10 on Sonnet.
Before & After Examples
What the extraction prompt produces when run on a real conversation.
3-month coding project chat
Before — buried in old chat
Preferences scattered across 200+ messages: "use TypeScript strict mode", "I hate default exports", "always write tests alongside code", "name variables after what they do not what they are"... all forgotten when starting a new project.
After — skill.md extracted
skill: coding-standards
trigger: Whenever writing or reviewing code
gotchas: User prefers strict mode, no any types
Preferences: Functional React components, named exports, tests in same PR, variables named by behavior not type...
trigger: Whenever writing or reviewing code
gotchas: User prefers strict mode, no any types
Preferences: Functional React components, named exports, tests in same PR, variables named by behavior not type...
Before — re-explaining every session
"Remember, I don't like bullet lists in my articles. Write in paragraphs. Keep sentences short. I always want a contrarian angle. Don't start with a statistic." — typed every new conversation.
After — installed once, always active
skill: content-voice
trigger: When writing articles or blog posts
Anti-Patterns: No bullet lists in articles, no stat openers, no passive voice, no "In conclusion..."
Style: Paragraph prose, short sentences, contrarian angle by default.
trigger: When writing articles or blog posts
Anti-Patterns: No bullet lists in articles, no stat openers, no passive voice, no "In conclusion..."
Style: Paragraph prose, short sentences, contrarian angle by default.
Model Compatibility
✓ Works Best With
- Claude Sonnet 4.6 — best at reading implicit preferences between the lines
- Claude Haiku 4.5 — efficient for shorter chats under 50 messages
- GPT-5.4 — strong alternative with wide context window
✗ Not Recommended
- GPT-5.4 Mini / Gemini Flash — truncate long conversation context
- Models without 100K+ context window (miss early conversation signals)
- Small local models — weak at inferring implicit patterns
Chainable Skills
The Memory Hack is the starting point for building a personal skill library. Run it on your old chats, then feed the output into the Specialist Stack or Orchestrator.
Old Chat
Months of buried context
→
Memory Hack
Extract · This skill
→
Skill.md File
Permanent reusable skill
→
Any Project
Drop in and it works
Pro workflow: Run Memory Hack on 3–5 of your most active chats, then feed the resulting skill.md files into the Specialist Stack. You'll end up with a curated folder of skills tailored to how you actually work — not generic prompts.
The System Prompt
Paste this as the system prompt, then send your entire old conversation as the user message. Claude will mine it and produce a complete skill.md.
Send the whole conversation. If you only send a summary, you get surface-level preferences. The real value is in the implicit patterns — what was changed, rejected, or quietly accepted — which only appear in the full context.
System Prompt
You are a skill architect. Your task is to analyze the entire conversation above and extract a reusable Claude skill from it. Look for: 1. PREFERENCES — what the user explicitly liked, approved, or asked for more of 2. REJECTIONS — what was changed, pushed back on, or discarded 3. WORKFLOW PATTERNS — recurring steps, processes, or sequences the user follows 4. STYLE SIGNALS — vocabulary, tone, formatting preferences, level of detail 5. DOMAIN KNOWLEDGE — technical context, jargon, constraints specific to this user 6. IMPLICIT RULES — things that were never stated but are clearly expected Output a complete skill.md file in this exact format: --- skill: [descriptive-slug] trigger: [when should Claude automatically invoke this skill — be specific] gotchas: [what commonly goes wrong without this skill] --- # [Skill Name] ## Purpose [One sentence: what this skill makes Claude do differently] ## Preferences Captured [Bulleted list of preferences extracted from the conversation] ## Anti-Patterns to Avoid [Bulleted list of things this user explicitly rejected] ## Workflow [Step-by-step process this user prefers for this type of work] ## Style Guide [Tone, formatting, vocabulary, level of detail] ## Context [Domain knowledge, constraints, background that Claude should always have] Output the skill.md only. No commentary. No explanation.
What Gets Extracted
The prompt targets six signal categories that accumulate naturally in any working conversation.
- Explicit preferences — "I like this", "keep this format", approval signals
- Rejections — everything you changed, deleted, or pushed back on in past sessions
- Workflow patterns — the recurring sequence you follow for this type of work
- Style signals — sentence length, vocabulary level, formatting choices, tone
- Domain knowledge — technical context, jargon, and constraints Claude needs to know
- Implicit rules — the unstated expectations that only show up as corrections when violated
The implicit rules are the real gold. Stated preferences are easy to remember. Implicit expectations — the ones you only notice when Claude violates them — are impossible to recall months later. This skill captures them before they're forgotten.
Community Signals
893
Total installs
4.7★
Community rating
3.2
Avg chats mined
Most commonly chained with: Specialist Stack (71% of workflows), Orchestrator (44%), Voice Fingerprint (38%). Most users run this on 2–5 old chats in their first session and build a personal skill folder from the results.
Ready to recover what you've already taught Claude?
Copy the prompt, open an old conversation, paste the whole thing as your message. That's it.
✓ Prompt copied to clipboard