AgentSkillsCN

learn

从当前会话中提取可复用的模式,并将其保存为直觉。

SKILL.md
--- frontmatter
name: learn
description: Extract reusable patterns from the current session and save them as instincts

Learn Skill

You are analyzing the current session to extract reusable patterns (instincts) that can improve future sessions.

What is an Instinct?

An instinct is a small, atomic learned behavior:

  • One trigger, one action - Specific and focused
  • Confidence-weighted - Score from 0.3 to 0.9
  • Evidence-backed - Documented with examples
  • Domain-tagged - Categorized for retrieval

Patterns to Detect

Look for these types of learnable patterns:

  1. Error Resolution - How specific errors were debugged and fixed
  2. User Corrections - When the user corrected your approach
  3. Workarounds - Solutions for limitations or edge cases
  4. Debugging Techniques - Effective debugging strategies
  5. Project Conventions - Codebase-specific patterns
  6. Workflow Patterns - Efficient sequences of operations
  7. Tool Preferences - Preferred tools or approaches

Patterns to Ignore

Skip these (not worth learning):

  • Simple typos
  • One-time fixes unlikely to recur
  • External API issues
  • Trivial operations

Instructions

  1. Review the session: Analyze the conversation for non-trivial patterns

  2. Identify candidates: List potential instincts with:

    • Trigger: What situation activates this?
    • Action: What should be done?
    • Evidence: What happened in this session?
    • Confidence: How reliable is this pattern? (0.3-0.9)
    • Domain: What category? (error, workflow, convention, etc.)
  3. Create instinct files: For each approved pattern, create a file in .claude/learning/instincts/:

markdown
# Instinct: [name]

## Trigger
[When this instinct applies]

## Action
[What to do when triggered]

## Evidence
- Session: [date]
- Context: [what happened]

## Metadata
- Confidence: [0.3-0.9]
- Domain: [category]
- Created: [timestamp]
  1. Report: Summarize what was learned

Confidence Scoring

  • 0.3-0.4: Observed once, might be situational
  • 0.5-0.6: Observed pattern, likely useful
  • 0.7-0.8: Strong pattern, auto-approve threshold
  • 0.9: Highly reliable, verified multiple times

Storage Location

Instincts are stored in .claude/learning/instincts/

Output Format

code
## Learning Summary

### New Instincts Created
1. [instinct-name] (confidence: X.X, domain: Y)
   - Trigger: ...
   - Action: ...

### Patterns Skipped
- [reason]: [pattern description]

### Recommendations
- [suggestions for /evolve if patterns are clustering]