skill-extractor（技能提取器）

Skill Extractor

Extracts reusable knowledge from work sessions and saves it as a Claude Code skill.

When to Use

• Just solved a non-obvious problem through investigation
• Discovered a workaround that required trial-and-error
• Found a debugging technique that would help in similar situations
• Learned a project-specific pattern worth preserving
• Fixed an error where the root cause wasn't immediately apparent

When NOT to Use

• Simple documentation lookups (just bookmark the docs)
• Trivial fixes (typos, obvious errors)
• One-off project-specific configurations
• Knowledge that's already well-documented elsewhere
• Unverified solutions (wait until it actually works)

Finding Extraction Candidates

Use these prompts to identify knowledge worth extracting:

• "What did I just learn that wasn't obvious before starting?"
• "If I faced this exact problem again, what would I wish I knew?"
• "What error message or symptom led me here, and what was the actual cause?"
• "Is this pattern specific to this project, or would it help in similar projects?"
• "What would I tell a colleague who hits this same issue?"

If you can't answer at least two of these with something non-trivial, it's probably not worth extracting.

Command

/skill-extractor [--project] [context hint]

• Default: saves to ~/.claude/skills/[name]/SKILL.md
• --project: saves to .claude/skills/[name]/SKILL.md
• Context hint helps focus extraction (e.g., /skill-extractor the cyclic data DoS fix)

Extraction Process

Step 0: Check for Existing Skills

Before creating a new skill, search for existing ones that might cover the same ground:

# Check user skills
ls ~/.claude/skills/

# Check project skills
ls .claude/skills/

# Search by keyword
grep -r "keyword" ~/.claude/skills/ .claude/skills/ 2>/dev/null

If a related skill exists, consider updating it instead of creating a new one. See skill-lifecycle.md for guidance on when to update vs create.

Step 1: Identify the Learning

If $ARGUMENTS contains a context hint (e.g., "the cyclic data DoS fix"), use it to focus the extraction on that specific topic.

Analyze the conversation to identify:

• What problem was solved?
• What made the solution non-obvious?
• What would someone need to know to solve this faster next time?
• What are the exact trigger conditions (error messages, symptoms)?

Present a brief summary to the user:

I identified this potential skill:

**Problem:** [Brief description]
**Key insight:** [What made it non-obvious]
**Triggers:** [Error messages or symptoms]

Step 2: Quality Assessment

Evaluate the candidate skill against these criteria:

Present assessment to user and ask: "Proceed with extraction? [yes/no]"

The user decides whether to proceed regardless of how many criteria pass. Respect their judgment - if they say yes, extract; if no, skip.

Step 3: Gather Details

Ask the user:

1. Skill name - Suggest a kebab-case name based on context, let them override
2. Scope - User-level (default) or project-level (--project)

Step 4: Optional Research

If the topic involves a specific library or framework:

• Use web search to find current best practices
• Use Context7 MCP (if available) for official documentation
• Include relevant sources in the References section

Skip research for:

• Project-specific internal patterns
• Generic programming concepts
• Time-sensitive extractions

Step 5: Generate the Skill

Use the template from skill-template.md.

Quality standards: Follow quality-guide.md to ensure the skill provides lasting value. Key points:

• Behavioral guidance over reference dumps
• Explain WHY, not just WHAT
• Specific triggers that compete well against other skills

Step 6: Validate Before Saving

Run through the validation checklist in skill-template.md. If validation fails, fix the issues before saving.

Step 7: Save the Skill

Create the directory and save:

• User-level: ~/.claude/skills/[name]/SKILL.md
• Project-level: .claude/skills/[name]/SKILL.md

Report success:

Skill saved to: [path]

The skill will be available in future sessions when the context matches:
"[first line of description]"

Memory Consolidation

When extracting, consider how the new knowledge relates to existing skills:

Combine or separate?

• Combine if the new knowledge is a variation or edge case of an existing skill
• Separate if it has distinct trigger conditions or solves a fundamentally different problem
• When in doubt, start separate - you can always merge later

Update vs create:

• Update an existing skill when you've discovered additional edge cases, better solutions, or corrections
• Create a new skill when the knowledge has different trigger conditions, even if the domain is related

Cross-referencing:

• If skills are related but separate, add a "See also" section linking them
• Example: A skill for "debugging connection pool exhaustion" might link to "serverless cold start optimization"

Skill Lifecycle

Skills aren't permanent. See skill-lifecycle.md for guidance on:

• Updating skills with new discoveries
• Deprecating skills when tools or patterns change
• Archiving skills that are no longer relevant

Rationalizations to Reject

If you catch yourself thinking any of these, do NOT extract:

• "This might be useful someday" - Only extract verified, reusable knowledge
• "Let me just save everything" - Quality over quantity
• "The user didn't confirm but it seems valuable" - Always get explicit confirmation
• "I'll skip the 'When NOT to Use' section" - It's mandatory for good skills
• "The description can be vague" - Specific triggers are essential for discovery

Example Extraction

Scenario: User discovered that an AST visitor crashes with RecursionError when analyzing serialized files containing cyclic references (e.g., a list that contains itself).

Identified learning:

• Cyclic data structures create cyclic ASTs
• Visitor pattern without cycle tracking causes infinite recursion
• Need to track visited nodes or enforce depth limits

Generated skill name: cyclic-ast-visitor-hardening

Key sections:

• When to Use: "RecursionError in AST visitor", "analyzing untrusted serialized input"
• When NOT to Use: "Recursion from deeply nested (but acyclic) structures"
• Problem: Visitor doesn't track visited nodes, enters infinite loop on cycles
• Solution: Add visited: set parameter, check before recursing
• Verification: Cyclic test case completes without RecursionError