Continuous Learning

1---2name: continuous-learning3description: Automatically extract reusable patterns from Claude Code sessions and save them as learned skills for future use.4origin: ECC5---6 7# Continuous Learning Skill8 9Automatically evaluates Claude Code sessions on end to extract reusable patterns that can be saved as learned skills.10 11## When to Activate12 13- Setting up automatic pattern extraction from Claude Code sessions14- Configuring the Stop hook for session evaluation15- Reviewing or curating learned skills in `~/.claude/skills/learned/`16- Adjusting extraction thresholds or pattern categories17- Comparing v1 (this) vs v2 (instinct-based) approaches18 19## Status20 21This v1 skill is still supported, but `continuous-learning-v2` is the preferred path for new installs. Keep v1 when you explicitly want the simpler Stop-hook extraction flow or need compatibility with older learned-skill workflows.22 23## How It Works24 25This skill runs as a **Stop hook** at the end of each session:26 271. **Session Evaluation**: Checks if session has enough messages (default: 10+)282. **Pattern Detection**: Identifies extractable patterns from the session293. **Skill Extraction**: Saves useful patterns to `~/.claude/skills/learned/`30 31## Configuration32 33Edit `config.json` to customize:34 35```json36{37  "min_session_length": 10,38  "extraction_threshold": "medium",39  "auto_approve": false,40  "learned_skills_path": "~/.claude/skills/learned/",41  "patterns_to_detect": [42    "error_resolution",43    "user_corrections",44    "workarounds",45    "debugging_techniques",46    "project_specific"47  ],48  "ignore_patterns": [49    "simple_typos",50    "one_time_fixes",51    "external_api_issues"52  ]53}54```55 56## Pattern Types57 58| Pattern | Description |59|---------|-------------|60| `error_resolution` | How specific errors were resolved |61| `user_corrections` | Patterns from user corrections |62| `workarounds` | Solutions to framework/library quirks |63| `debugging_techniques` | Effective debugging approaches |64| `project_specific` | Project-specific conventions |65 66## Hook Setup67 68Add to your `~/.claude/settings.json`:69 70```json71{72  "hooks": {73    "Stop": [{74      "matcher": "*",75      "hooks": [{76        "type": "command",77        "command": "~/.claude/skills/continuous-learning/evaluate-session.sh"78      }]79    }]80  }81}82```83 84## Why Stop Hook?85 86- **Lightweight**: Runs once at session end87- **Non-blocking**: Doesn't add latency to every message88- **Complete context**: Has access to full session transcript89 90## Related91 92- [The Longform Guide](https://x.com/affaanmustafa/status/2014040193557471352) - Section on continuous learning93- `/learn` command - Manual pattern extraction mid-session94 95---96 97## Comparison Notes (Research: Jan 2025)98 99### vs Homunculus100 101Homunculus v2 takes a more sophisticated approach:102 103| Feature | Our Approach | Homunculus v2 |104|---------|--------------|---------------|105| Observation | Stop hook (end of session) | PreToolUse/PostToolUse hooks (100% reliable) |106| Analysis | Main context | Background agent (Haiku) |107| Granularity | Full skills | Atomic "instincts" |108| Confidence | None | 0.3-0.9 weighted |109| Evolution | Direct to skill | Instincts → cluster → skill/command/agent |110| Sharing | None | Export/import instincts |111 112**Key insight from homunculus:**113> "v1 relied on skills to observe. Skills are probabilistic—they fire ~50-80% of the time. v2 uses hooks for observation (100% reliable) and instincts as the atomic unit of learned behavior."114 115### Potential v2 Enhancements116 1171. **Instinct-based learning** - Smaller, atomic behaviors with confidence scoring1182. **Background observer** - Haiku agent analyzing in parallel1193. **Confidence decay** - Instincts lose confidence if contradicted1204. **Domain tagging** - code-style, testing, git, debugging, etc.1215. **Evolution path** - Cluster related instincts into skills/commands122 123See: `docs/continuous-learning-v2-spec.md` for full spec.