Pattern extractor

Pattern Extractor

Purpose

Close the learning loop by automatically analyzing transcript history to extract patterns, update technical knowledge, and suggest preventive measures.

Goal

Transform experience into actionable knowledge without manual intervention.

Use Cases

1. Automated Learning: Extract lessons from recent work
2. Error Prevention: Find repeated mistakes and suggest pre-flight checks
3. Best Practices: Identify successful patterns to replicate
4. Decision Documentation: Capture decision rationale for future reference
5. Cross-Agent Knowledge: Generate insights for sharing with other agents

When to Use

Trigger automatically:

• Daily (via cron): Extract patterns from last 24 hours
• Weekly: Extract patterns from last 7 days
• After major work: Extract patterns from specific session

Trigger manually:

• After encountering repeated errors
• Before starting similar work (learn from past attempts)
• When updating technical-patterns.md

Usage

# Extract patterns from recent transcripts
/pattern-extractor --last 7

# Extract specific pattern types
/pattern-extractor --errors-only --last 30
/pattern-extractor --decisions-only --since "2026-02-01"
/pattern-extractor --successes-only --last 14

# Update technical patterns file
/pattern-extractor --last 30 --update-file memory/technical-patterns.md

# Generate pre-flight checks from errors
/pattern-extractor --generate-checks --last 90

# Export for sharing
/pattern-extractor --last 7 --format json --output patterns.json

Process

1. Transcript Analysis

• Load transcripts from specified time range (uses transcript-query)
• Parse tool uses, errors, user feedback, outcomes
• Build timeline of activities

2. Pattern Detection

Error Patterns:

Analyze for:
- Tool errors (Read failed, Edit failed, Bash failed)
- Repeated mistakes (same error multiple times)
- Error sequences (Error A → leads to → Error B)
- Root causes (what triggered the error chain)

Success Patterns:

Analyze for:
- Successful completions (task done, tests pass, user approves)
- Techniques that worked (tool sequences, approaches)
- Fast resolutions (problem → solution quickly)
- User praise (positive feedback patterns)

Decision Patterns:

Analyze for:
- Architecture decisions (chose X over Y because Z)
- Trade-offs considered (pros/cons lists)
- Rationale documented (why this approach)
- Alternatives rejected (what was considered but not chosen)

3. Pattern Categorization

Anti-Patterns (things to avoid):

• Frequency: How often does this mistake occur?
• Impact: What's the cost (time lost, bugs introduced)?
• Root cause: Why does this happen?
• Prevention: How to stop it?

Best Practices (things to replicate):

• Success rate: How reliably does this work?
• Context: When is this applicable?
• Steps: What's the procedure?
• Metrics: How to measure success?

Decisions (for future reference):

• Problem: What was being solved?
• Solution: What was chosen?
• Reasoning: Why was this the best option?
• Outcome: Did it work?

4. Knowledge Base Update

Update technical-patterns.md:

## Anti-Patterns (Last Updated: 2026-02-06)

### Pattern: Always Read before Write/Edit
**Frequency**: 22 occurrences (last 90 days)
**Impact**: File operations fail, wasted time
**Root Cause**: Assumption file exists without verification
**Prevention**:
- Pre-flight check: Has this file been read in session?
- Auto-suggest: "File not read yet - would you like to read it first?"
**Last Seen**: 2026-02-05 (session abc123...)

### Pattern: Verify paths when switching repos
**Frequency**: 21 occurrences (last 90 days)
**Impact**: Operations on wrong files, potential data loss
**Root Cause**: Context switching without path verification
**Prevention**:
- Pre-flight check: Verify cwd matches expected repo
- Prompt pattern: "Working in {repo} - is this correct?"
**Last Seen**: 2026-02-04 (session def456...)

Update best-practices.md (if exists):

## Best Practices (Last Updated: 2026-02-06)

### Pattern: Use native FormData for file uploads
**Success Rate**: 100% (5/5 attempts)
**Context**: Uploading files to APIs requiring multipart/form-data
**Steps**:
1. Create FormData instance
2. Append file as Blob (not Buffer)
3. Let fetch auto-set Content-Type boundary
4. Don't manually set Content-Type header
**Metrics**: Zero upload failures after adopting this pattern
**First Success**: 2026-02-05

5. Generate Preventive Measures

Pre-flight Checks (code suggestions):

// Suggested pre-flight check for Write/Edit tools
function preWriteCheck(filePath, sessionContext) {
  const hasBeenRead = sessionContext.filesRead.includes(filePath);

  if (!hasBeenRead) {
    return {
      warning: true,
      message: `File ${filePath} has not been read in this session. Read it first?`,
      suggestion: `Read(${filePath})`
    };
  }

  return { warning: false };
}

Tool Wrappers (safety layer):

// Suggested wrapper for Bash tool when switching repos
async function safeBash(command, expectedRepo) {
  const cwd = process.cwd();

  if (!cwd.includes(expectedRepo)) {
    const userConfirm = await askUser(
      `Command will run in ${cwd}. Expected ${expectedRepo}. Continue?`
    );
    if (!userConfirm) return { cancelled: true };
  }

  return await Bash(command);
}

6. Output Formats

Markdown (human-readable, for memory files):

# Pattern Analysis: 2026-02-06

## Summary
- Transcripts analyzed: 42 (last 30 days)
- Error patterns found: 8
- Success patterns found: 12
- Decisions documented: 6

## New Insights
...

JSON (machine-readable, for sharing):

{
  "generated": "2026-02-06T14:30:00Z",
  "period": {
    "days": 30,
    "transcripts": 42,
    "sessions": 38
  },
  "patterns": {
    "errors": [
      {
        "pattern": "Read before Write violation",
        "frequency": 22,
        "impact": "high",
        "prevention": "pre-flight check",
        "code": "function preWriteCheck() { ... }"
      }
    ],
    "successes": [...],
    "decisions": [...]
  }
}

Integration Points

1. Daily Automation (cron)

# Run daily at 6am, update technical patterns
0 6 * * * cd <project-root> && /pattern-extractor --last 1 --update-file memory/technical-patterns.md

2. Session Start Hook

// Load recent patterns at session start
import { exec } from 'child_process';

const patterns = await exec('pattern-extractor --last 7 --format json');
console.log(`Recent patterns: ${patterns.errors.length} errors, ${patterns.successes.length} wins`);

3. Pre-Tool Execution Hooks

// Before Write tool
const writeCheck = await exec(`pattern-extractor --check Write --file ${filePath}`);
if (writeCheck.warnings.length > 0) {
  // Show warnings to user
}

4. Cross-Agent Knowledge Sharing

// Extract patterns and broadcast to other agents
const patterns = await exec('pattern-extractor --last 7 --format json');

await agentDispatch.broadcast({
  type: 'knowledge_share',
  from: '<agent-id>',
  patterns: patterns,
  timestamp: new Date().toISOString()
});

Architecture

Dependencies

• transcript-query skill (transcript parsing)
• Node.js fs/promises (file I/O)
• Optional: LLM API (for pattern summarization)

File Structure

.claude/skills/pattern-extractor/
├── SKILL.md                   # This file
├── pattern-extractor.mjs      # CLI entry point
├── lib/
│   ├── pattern-detector.js    # Core pattern detection logic
│   ├── error-analyzer.js      # Error pattern analysis
│   ├── success-analyzer.js    # Success pattern analysis
│   ├── decision-analyzer.js   # Decision pattern analysis
│   └── check-generator.js     # Pre-flight check code generation
└── templates/
    ├── technical-patterns.md  # Template for pattern doc
    └── pre-flight-checks.js   # Template for checks

Key Algorithms

Pattern Detection Algorithm

For each transcript in time range:
  1. Parse tool uses and outcomes
  2. Identify error sequences
  3. Detect repeated patterns (same error multiple times)
  4. Find successful completions
  5. Extract decision points

Group similar patterns:
  - Cluster by error message similarity
  - Cluster by tool sequence similarity
  - Cluster by decision context

Rank by importance:
  - Frequency × Impact = Priority
  - High priority patterns bubble up

Prevention Suggestion Algorithm

For each error pattern:
  1. Identify trigger condition
  2. Determine detection point (when can we catch this?)
  3. Generate pre-flight check code
  4. Suggest integration point (which tool hook?)

Example:
  Error: "Write failed - file not read"
  Trigger: Write tool called without prior Read
  Detection: Before Write execution
  Check: Has file been read in session?
  Integration: Pre-Write hook

Output Example

# Pattern Analysis Report
**Generated**: 2026-02-06 14:30 UTC
**Period**: Last 30 days (42 transcripts)

---

## High-Priority Anti-Patterns

### 1. Read-Before-Write Violation
**Occurrences**: 22 times (0.73/day)
**Impact**: High (causes file operation failures)
**Cost**: ~15 min/occurrence = 5.5 hours wasted

**Pattern**:

Write(file.ts) → Error: "File not read" → Read(file.ts) → Write(file.ts) → Success

**Root Cause**: Assumption that file context is available without reading

**Prevention**:
```javascript
// Suggested pre-Write hook
if (!session.filesRead.includes(targetFile)) {
  warn("File not read. Reading automatically...");
  await Read(targetFile);
}

Recommendation: Implement pre-flight check — HIGH PRIORITY

---

Successful Patterns to Replicate

1. Native FormData for Uploads

Success Rate: 100% (5/5) Context: File uploads to external APIs

Pattern:

const formData = new FormData();
formData.append('file', new Blob([fileBuffer]), filename);
// Let fetch auto-set Content-Type - don't set manually
const response = await fetch(url, { method: 'POST', body: formData });

Why it works: APIs need proper boundary in Content-Type for multipart uploads

When to use: Any external API requiring multipart/form-data

---

Recent Decisions

1. Agent Teams Integration in dev-workflow-orchestrator

Date: 2026-02-06 Context: Should workflow offer agent teams automatically? Decision: Yes, with analysis phase Reasoning:

• Intelligent automation - offers when beneficial
• User maintains control - can decline
• Clear cost/benefit - shows 3-5x token cost vs 2-3x speed Alternatives Considered:
• Always use teams (too expensive) — rejected
• Never suggest teams (misses opportunities) — rejected
• Analyze and suggest (balanced approach) — chosen Outcome: Implemented in phase 3 of workflow

---

Recommendations

1. Implement pre-flight checks (code provided above) - prevents 22 errors/month
2. Replicate FormData pattern - 100% success rate for uploads
3. Document decision pattern - agent teams suggestion logic is reusable

---

Next Pattern Extraction: 2026-02-07 06:00 UTC (automated)

## Limitations

- Pattern detection requires clean transcript data
- Code generation is suggestions only (needs review)
- Cross-agent patterns need standardized format
- May miss subtle patterns that humans would catch

## Future Enhancements

1. **ML-based pattern detection**: Use embeddings to find semantic patterns
2. **Auto-fix suggestions**: Generate full PR for common fixes
3. **Pattern prediction**: "You're about to make error X based on context"
4. **Cross-repo patterns**: Find patterns across multiple projects
5. **User feedback loop**: Track if suggested patterns actually help

## Security & Privacy

- Patterns may contain sensitive information from transcripts
- Filter out credentials, API keys, personal data
- Sanitize before sharing
- Respect transcript privacy settings

---

**Created**: 2026-02-06
**Dependencies**: transcript-query, Node.js 18+
**Related Skills**: brain-briefing, transcript-query, memory-manager
**Status**: Ready to implement