Continuous Learning Skill
Purpose
Automatically evaluate sessions and extract reusable patterns that improve future sessions.
How It Works
Runs as a Stop hook at session end:
- Checks if session has enough activity (default: 10+ tool calls)
- Identifies extractable patterns (error resolutions, user corrections, workarounds)
- Saves patterns to
~/.claude/homunculus/instincts/personal/
Configuration
config.json in this skill directory:
{
"min_session_length": 10,
"extraction_threshold": "medium",
"patterns_to_detect": [
"error_resolution",
"user_corrections",
"workarounds",
"debugging_techniques",
"project_specific"
],
"ignore_patterns": [
"simple_typos",
"one_time_fixes",
"external_api_issues"
]
}
Pattern Types
| Pattern | Description |
|---|---|
error_resolution |
How specific errors were resolved |
user_corrections |
Patterns from user corrections |
workarounds |
Solutions to framework quirks |
debugging_techniques |
Effective debugging approaches |
project_specific |
Project-specific conventions |
Integration with Auto-Learning
This skill feeds into the existing auto-learning system:
- Extracted patterns become instincts in
~/.claude/homunculus/instincts/personal/ - High-confidence instincts can evolve into skills via
/idev:evolve
Anti-Patterns
- Do NOT extract trivial patterns (typos, one-off fixes)
- Do NOT extract patterns without evidence (must be observed 2+ times)
- Do NOT overwrite existing instincts — update confidence instead