Extraction Recommender

Identifies patterns that should be extracted into reusable components.

Philosophy

"If you do it 3 times, script it. If it needs reasoning, agentize it."

Detect repeated patterns and recommend appropriate extraction:

• Script: Deterministic, repeatable operations
• Agent: LLM-required reasoning/analysis
• Skill: Reusable knowledge/guidance

---

Extraction Decision Tree

Is this pattern repeated 3+ times?
├─ No → Don't extract (premature abstraction)
└─ Yes → Does it require LLM reasoning?
    ├─ Yes → Is it a workflow (multi-step)?
    │   ├─ Yes → Extract as AGENT
    │   └─ No → Is it knowledge/guidance?
    │       ├─ Yes → Extract as SKILL
    │       └─ No → Extract as AGENT (single-task)
    └─ No → Is it deterministic?
        ├─ Yes → Extract as SCRIPT
        └─ No → Is it configuration-driven?
            ├─ Yes → Extract as TEMPLATE
            └─ No → Manual case-by-case

---

Detection Patterns

Script Extraction Candidates

Indicators:
  - Same bash commands repeated in multiple places
  - File manipulation with consistent pattern
  - Validation logic duplicated
  - Build/test steps repeated

Examples:
  - "rm -rf __pycache__ && rm -rf .pytest_cache"
  - "python3 scripts/validate_all.py --json"
  - "git add . && git commit -m ..."

Detection:
  - Grep for repeated command sequences
  - Look for copy-paste patterns
  - Check for hardcoded paths repeated

Agent Extraction Candidates

Indicators:
  - Complex analysis requiring reasoning
  - Multi-step decision making
  - Code review/quality assessment
  - Architectural analysis

Examples:
  - "Analyze this PR for issues"
  - "Review code for security vulnerabilities"
  - "Determine appropriate hook type"

Detection:
  - Look for analysis-heavy instructions
  - Find multi-step reasoning patterns
  - Check for quality judgment tasks

Skill Extraction Candidates

Indicators:
  - Referenced knowledge blocks
  - Repeated prompt patterns
  - Domain-specific guidance
  - Best practices documentation

Examples:
  - "When working with MCP, always..."
  - "The correct pattern for hooks is..."
  - "Progressive disclosure means..."

Detection:
  - Look for repeated "best practices" text
  - Find duplicated guidance sections
  - Check for domain knowledge blocks

---

Detection Process

Phase 1: Pattern Discovery

Scan for repeated patterns:

# Conceptual detection logic
patterns_found = []

# 1. Command sequences
bash_patterns = find_repeated_commands(threshold=3)

# 2. Analysis instructions
analysis_patterns = find_reasoning_tasks(threshold=2)

# 3. Knowledge blocks
knowledge_patterns = find_guidance_text(threshold=2)

patterns_found = bash_patterns + analysis_patterns + knowledge_patterns

Phase 2: Classification

For each pattern found:

Question	Yes →	No →
Requires LLM?	Agent/Skill	Script
Multi-step workflow?	Agent	Skill (if guidance)
Deterministic?	Script	Agent
Configuration-driven?	Template	Case-by-case

Phase 3: Generate Recommendations

For each extraction candidate:

1. Identify pattern with examples
2. Classify type (script/agent/skill)
3. Generate template for extraction
4. Estimate effort (low/medium/high)

---

Output Format

## Extraction Analysis

**Files Scanned:** {count}
**Patterns Found:** {count}
**Extraction Candidates:** {count}

### Summary

| Type | Candidates | Estimated Effort |
|------|------------|------------------|
| Script | {n} | {total hours} |
| Agent | {n} | {total hours} |
| Skill | {n} | {total hours} |

### Script Extraction Candidates

#### Candidate: {name}

**Pattern Found:**
\`\`\`bash
{repeated_command_sequence}
\`\`\`

**Occurrences:** {count} times in {files}

**Recommended Extraction:**
\`\`\`bash
#!/bin/bash
# scripts/{name}.sh
# Extracted from: {source_files}

{generated_script}
\`\`\`

**Usage After Extraction:**
\`\`\`bash
bash scripts/{name}.sh
\`\`\`

**Effort:** {low|medium|high}

---

### Agent Extraction Candidates

#### Candidate: {name}

**Pattern Found:**
{analysis_task_description}

**Occurrences:** {count} times

**Recommended Extraction:**
\`\`\`yaml
# agents/{name}.md
---
name: {name}
description: {extracted_from_pattern}
tools: [{recommended_tools}]
model: {recommended_model}
---

{generated_agent_structure}
\`\`\`

**Effort:** {low|medium|high}

---

### Skill Extraction Candidates

#### Candidate: {name}

**Pattern Found:**
{knowledge_block}

**Occurrences:** {count} times in {files}

**Recommended Extraction:**
\`\`\`yaml
# skills/{name}/SKILL.md
---
name: {name}
description: {extracted_description}
allowed-tools: [{recommended_tools}]
---

{generated_skill_structure}
\`\`\`

**Effort:** {low|medium|high}

---

### Not Recommended for Extraction

| Pattern | Reason | Occurrences |
|---------|--------|-------------|
| {pattern} | Only {n} times (<3) | {n} |
| {pattern} | Too context-specific | {n} |

### Prioritized Actions

| Priority | Type | Pattern | Action |
|----------|------|---------|--------|
| HIGH | Script | {pattern} | Extract to scripts/{name}.sh |
| MEDIUM | Agent | {pattern} | Create agents/{name}.md |
| LOW | Skill | {pattern} | Create skills/{name}/ |

---

Extraction Templates

Script Template

#!/bin/bash
# {name}.sh - Extracted pattern
#
# Source: {original_locations}
# Purpose: {what_it_does}
# Usage: bash scripts/{name}.sh [args]

set -e

# Configuration
{extracted_config}

# Main logic
{extracted_logic}

# Output
echo "{completion_message}"

Agent Template

---
name: {name}
description: {purpose}
tools: [{tools}]
skills: {related_skills}
model: {model}
---

# {Name}

{brief_description}

## Process

### Phase 1: {first_step}
{extracted_logic}

### Phase 2: {second_step}
{extracted_logic}

## Output Format
{format}

## Success Criteria
{criteria}

Skill Template

---
name: {name}
description: {purpose}
allowed-tools: [{tools}]
---

# {Name}

{brief_description}

## When to Use

- {trigger_1}
- {trigger_2}

## How to Use

{extracted_guidance}

## References

- [Detail 1](references/{ref1}.md)

---

Success Criteria

• All relevant files scanned
• Repeated patterns identified (threshold: 3+)
• Patterns classified correctly (script/agent/skill)
• Templates generated for each candidate
• Effort estimated
• Priorities assigned
• Non-candidates explained