jikime-workflow-learning

name: jikime-workflow-learning description: Continuous learning system - extract, store, and reuse patterns from Claude Code sessions version: 1.0.0 tags: ["workflow", "learning", "patterns", "knowledge", "session", "improvement"] triggers: keywords: ["learn", "pattern", "lesson", "remember", "extract", "학습"] phases: ["sync"] agents: ["manager-docs", "manager-strategy"] languages: [] progressive_disclosure: enabled: true level1_tokens: ~100 level2_tokens: ~7000 user-invocable: false context: fork agent: manager-docs allowed-tools: - Read - Write - Edit - Bash - Grep - Glob - TodoWrite

Continuous Learning Skill

Automatically extract reusable patterns from Claude Code sessions and store them for future use.

Philosophy

Every session is a learning opportunity:
├─ Error resolutions → Future prevention
├─ User corrections → Preference learning
├─ Workarounds → Knowledge base
├─ Debugging techniques → Reusable strategies
└─ Project-specific patterns → Team knowledge

How It Works

Session Lifecycle

Session Start
    ↓
  Load relevant learnings from .jikime/learnings/
    ↓
  Development work...
    ↓
Session End (Stop hook)
    ↓
  Analyze session for patterns
    ↓
  Extract learnings with confidence scoring
    ↓
  Store in .jikime/learnings/
    ↓
Next Session: Patterns available

Automatic Extraction

At session end, the system analyzes:

• Error messages and resolutions
• User corrections to Claude's suggestions
• Workarounds for framework/library quirks
• Debugging techniques that worked
• Project-specific conventions

Pattern Categories

For detailed YAML examples and patterns, see:

• Pattern Categories - Error resolution, user corrections, workarounds, debugging, conventions

Storage Structure

.jikime/
├── learnings/
│   ├── index.json              # Searchable index
│   ├── errors/
│   │   ├── typescript.yaml
│   │   ├── react.yaml
│   │   └── nextjs.yaml
│   ├── corrections/
│   │   └── style-preferences.yaml
│   ├── workarounds/
│   │   ├── nextjs-14.yaml
│   │   └── prisma.yaml
│   ├── debugging/
│   │   └── react-state.yaml
│   ├── conventions/
│   │   ├── api-patterns.yaml
│   │   └── file-structure.yaml
│   └── sessions/
│       ├── 2024-01-22-summary.md
│       └── 2024-01-21-summary.md

Index Structure

{
  "version": "1.0.0",
  "last_updated": "2024-01-22T15:30:00Z",
  "total_patterns": 47,
  "categories": {
    "error_resolution": 15,
    "user_correction": 8,
    "workaround": 12,
    "debugging": 7,
    "project_convention": 5
  },
  "top_patterns": [
    {"id": "err-ts-001", "confidence": 0.95, "frequency": 12},
    {"id": "wk-nextjs-003", "confidence": 0.93, "frequency": 8}
  ],
  "technologies": ["typescript", "react", "nextjs", "prisma"]
}

Confidence Scoring

Patterns are scored for reliability:

Confidence = (
  base_score * 0.4 +
  frequency_score * 0.3 +
  recency_score * 0.2 +
  source_reliability * 0.1
)

Base Score:
  - Official docs: 1.0
  - User correction: 0.9
  - Successful resolution: 0.8
  - Experimental: 0.6

Frequency Score:
  - Used 10+ times: 1.0
  - Used 5-9 times: 0.8
  - Used 2-4 times: 0.6
  - Used once: 0.4

Recency Score:
  - Used this week: 1.0
  - Used this month: 0.8
  - Used this quarter: 0.6
  - Older: 0.4

Confidence Thresholds

Orchestrator Integration

J.A.R.V.I.S. (Development)

Session Start:
  → Load high-confidence patterns for active technologies
  → Summarize: "Loaded 12 patterns for React/TypeScript"

During Development:
  → Apply patterns proactively
  → "Based on learned pattern: using optional chaining here"

Session End:
  → Extract new patterns
  → Report: "3 new patterns learned this session"

Predictive Suggestions:
  → "Based on past sessions, you might also want to..."

F.R.I.D.A.Y. (Migration)

Migration Start:
  → Load patterns for source/target frameworks
  → "Loaded 8 migration patterns for Vue → React"

During Migration:
  → Apply migration-specific workarounds
  → Track framework-specific quirks

Migration End:
  → Store migration patterns for future use
  → Export as reusable migration guide

Session Summary

At session end, generate a summary:

# Session Summary: 2024-01-22

## Duration
Started: 10:30 AM
Ended: 2:15 PM (3h 45m)

## Work Completed
- Implemented user authentication
- Fixed 3 TypeScript errors
- Resolved hydration mismatch issue

## Patterns Learned

### New Patterns (3)
1. **Error Resolution**: TypeScript strict null checks
   - Confidence: 0.85
   - Category: error_resolution

2. **Workaround**: Next.js 14 cache invalidation
   - Confidence: 0.78
   - Category: workaround

3. **Convention**: API response structure
   - Confidence: 0.92
   - Category: project_convention

### Reinforced Patterns (2)
- React useState with objects (frequency: 5 → 6)
- Prisma relation queries (frequency: 3 → 4)

## For Next Session
- Continue with payment integration
- Review auth edge cases
- Consider adding rate limiting

Export/Import

For detailed CLI commands and export format, see:

• CLI Commands - Export, import, and search commands

Quick reference:

# Export patterns
jikime-adk learnings export --output learnings-export.yaml

# Import from another project
jikime-adk learnings import --source patterns.yaml --strategy merge

Hook Integration

Session End Hook

{
  "hooks": {
    "Stop": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "jikime-adk hooks learning-extract"
          }
        ]
      }
    ]
  }
}

Session Start Hook

{
  "hooks": {
    "SessionStart": [
      {
        "matcher": "*",
        "hooks": [
          {
            "type": "command",
            "command": "jikime-adk hooks learning-load"
          }
        ]
      }
    ]
  }
}

Configuration

# .jikime/config/learning.yaml

learning:
  enabled: true

  # Extraction settings
  extraction:
    min_session_length: 10          # Minimum messages to analyze
    auto_extract: true              # Extract on session end
    require_confirmation: false     # Ask before saving patterns

  # Pattern settings
  patterns:
    min_confidence: 0.40            # Minimum to store
    auto_apply_threshold: 0.85      # Apply without asking
    max_age_days: 365               # Archive old patterns

  # Categories to track
  categories:
    - error_resolution
    - user_correction
    - workaround
    - debugging
    - project_convention

  # Ignore patterns
  ignore:
    - simple_typos
    - one_time_fixes
    - external_api_issues
    - environment_specific

Searching Patterns

Query stored patterns (see CLI Commands for full examples):

# Search by keyword
jikime-adk learnings search "useState"

# Search by category
jikime-adk learnings search --category workaround

# Search by technology
jikime-adk learnings search --tech nextjs

Privacy & Security

Sensitive Data Handling

# Patterns never stored:
- API keys, tokens, secrets
- Passwords or credentials
- Personal information
- Environment-specific values

# Before storage:
- Redact secrets: sk-*** → [REDACTED]
- Generalize specific values
- Remove project-specific paths

Local Storage Only

All learnings stored locally in .jikime/learnings/
- Not synced to cloud by default
- Export explicitly for sharing
- Add to .gitignore if sensitive

Best Practices

DO

1. Review high-frequency patterns - They shape future behavior
2. Adjust confidence when wrong - Learning improves over time
3. Export valuable patterns - Share across projects
4. Clean stale patterns - Remove outdated learnings
5. Categorize correctly - Aids future retrieval

DON'T

1. Trust low-confidence blindly - Verify before applying
2. Store one-time fixes - Not reusable
3. Keep outdated patterns - Technology evolves
4. Ignore user corrections - They signal preferences
5. Over-generalize - Some patterns are context-specific

Works Well With

• jikime-foundation-core: Core workflow integration
• jikime-workflow-spec: SPEC-based development
• jikime-workflow-eval: Evaluation framework
• jikime-workflow-project: Project initialization
• jikime-foundation-claude: Claude Code patterns

Last Updated: 2026-01-25 Version: 1.0.0 Integration: SessionEnd hook, J.A.R.V.I.S./F.R.I.D.A.Y., Export/Import