Skip to main content

Agent Team Assembly Strategy for Prompt Library Generation

Task: Extract prompts from 151,000+ words of research and create categorized prompt library Created: 2025-11-29 Strategy: Multi-agent orchestration with specialized roles

🎯 Optimal Agent Team (7 Agents + 1 Orchestrator)​

Phase 1: Research Extraction & Pattern Discovery​

Agent 1: codebase-locator​

Role: Initial reconnaissance - find WHERE prompts live Tools: Grep, Glob, LS Why This Agent:

  • ✅ Specialized in finding locations - Designed specifically for "WHERE is X?" queries
  • ✅ Efficient search patterns - Uses optimized Grep/Glob strategies
  • ✅ Organized results - Returns structured file listings by category
  • ✅ Fast execution - Haiku model for speed, perfect for discovery phase

Specific Task:

Search all 9 research documents for:
- Sections containing "prompt", "template", "example"
- Code blocks with video generation syntax
- Sections titled "Prompt Template", "Prompting", "Examples"
- Return file:line references for all matches

Expected Output:

research-video-instructional-material-prompting-techniques.md:198-350 (Part 107 Prompt Templates)
research-video-instructional-material-prompting-techniques.md:651-800 (Master Style Reference)
google-veo-comprehensive-guide.md:859-975 (Veo Prompt Templates)
ai-video-generation-howto.md:450-600 (Tool-Specific Prompts)
...

Agent 2: codebase-pattern-finder​

Role: Extract concrete prompt examples and identify patterns Tools: Read, Write, Edit, Grep, Glob, LS, TodoWrite Why This Agent:

  • ✅ Pattern recognition specialist - Designed to find similar implementations
  • ✅ Provides code details - Not just locations, but actual prompt content
  • ✅ Auto-categorization - Can organize prompts by complexity (simple vs advanced)
  • ✅ Confidence indicators - Scores pattern quality and relevance

Specific Task:

Based on locations from codebase-locator:
1. Read each identified section
2. Extract all prompt examples (code blocks, templates, syntax patterns)
3. Identify prompt structure patterns (anatomy, components, formatting)
4. Group similar prompts (instructor-led, B-roll, animations, etc.)
5. Assess prompt quality and completeness

Expected Output:

{
  "instructor_prompts": [
    {
      "source": "RESEARCH...md:210-225",
      "pattern": "HeyGen avatar + professional tone",
      "confidence": 0.95,
      "example": "Medium shot of [instructor]...",
      "complexity": "basic"
    },
    ...
  ],
  "b_roll_prompts": [...],
  "animation_prompts": [...]
}

Agent 3: codebase-analyzer​

Role: Deep-dive technical analysis of prompt anatomy and best practices Tools: Read, Write, Edit, Grep, Glob, LS, TodoWrite Why This Agent:

  • ✅ Understands HOW things work - Analyzes implementation details
  • ✅ Architectural pattern recognition - Identifies prompt design patterns
  • ✅ Technical documentation - Can explain complex technical concepts
  • ✅ Integration analysis - Understands how prompts work with different tools (HeyGen, Veo, etc.)

Specific Task:

Analyze extracted prompts to understand:
1. Prompt anatomy (7-component structure from research)
2. Platform-specific requirements (Veo vs HeyGen vs Runway syntax)
3. Consistency techniques (Ingredients to Video, style references)
4. Quality factors (what makes a prompt effective)
5. Common mistakes and anti-patterns
6. Integration with production workflows

Expected Output:

## Prompt Anatomy Analysis

### Universal Structure (7 Components):
1. Shot Type - Determines framing (MS, CU, WS, etc.)
2. Subject - Primary focus (instructor, drone, diagram)
3. Action - Movement/behavior
4. Environment - Setting/background
5. Style - Visual aesthetic
6. Camera Movement - Motion (dolly, pan, static)
7. Audio Description - Sound requirements

### Platform-Specific Variations:
- **HeyGen**: Requires avatar selection, voice clone ID, minimal shot direction
- **Veo**: Supports cinematic terminology, duration specification, "Ingredients to Video"
- **Runway**: Creative freedom, abstract descriptions acceptable
...

Phase 2: Content Organization & Documentation​

Agent 4: documentation-librarian​

Role: Organize extracted prompts into category structure Tools: Read, Write, Edit, Bash, Glob, Grep, LS, TodoWrite Why This Agent:

  • ✅ Documentation organization specialist - Designed for creating navigable systems
  • ✅ Auto-indexing - Can generate indexes and navigation files automatically
  • ✅ Cross-reference management - Links related content across categories
  • ✅ Freshness monitoring - Ensures content stays up-to-date

Specific Task:

Organize extracted prompts into directory structure:

/prompts/
  01-educational-content/
    instructor-segments/
      - Categorize all instructor-related prompts here
      - Group by difficulty (basic, intermediate, advanced)
    concept-explanations/
    procedure-demonstrations/
    assessment-content/
  02-technical-training/
    aviation/ - PRIORITY for Part 107
    technology/
    professional/
  03-b-roll-footage/
  04-animations/
  05-specialized/

Create taxonomy and ensure no duplicate prompts across categories.

Expected Output:

✅ Prompt catalog organized (120+ prompts categorized)
✅ Cross-reference map created (related prompts linked)
✅ Duplicate detection complete (3 duplicates found, resolved)
✅ Priority ranking applied (aviation=high, general=medium, specialized=low)

Agent 5: codi-documentation-writer​

Role: Generate comprehensive README.md and CLAUDE.md for each category Tools: Read, Write, Edit, Grep, Glob, TodoWrite, Bash Why This Agent:

  • ✅ Comprehensive technical writer - Creates API docs, user guides, system documentation
  • ✅ Quality assurance expert - Ensures documentation meets standards
  • ✅ Enterprise-grade quality - 40/40 scoring, ADR compliance
  • ✅ Cross-document consistency - Maintains uniform style across all files

Specific Task:

For EACH of 17 directories (5 parent + 12 child):

1. README.md:
   - Category overview and purpose
   - When to use these prompts
   - Best practices specific to this category
   - Quick reference tables
   - Related categories
   - 5-10 prompt examples with explanations

2. CLAUDE.md:
   - AI agent automation guidance
   - Integration with CODITECT workflow
   - Quality standards for this prompt type
   - Common mistakes to avoid
   - Programmatic usage examples

Expected Output: 17 directories × 2 files = 34 documentation files (README + CLAUDE)

Phase 3: Prompt Template Generation​

Agent 6: educational-content-generator​

Role: Create prompt examples and templates at multiple skill levels Tools: Read, Write, Edit, Bash, Grep, Glob, TodoWrite Why This Agent:

  • ✅ Multi-level content creation - Generates beginner, intermediate, advanced versions
  • ✅ Pedagogical expertise - Applies learning science principles
  • ✅ NotebookLM optimization - Formats content for AI consumption
  • ✅ Progressive difficulty - Ensures appropriate complexity scaling

Specific Task:

For each of 12 subcategories, create:

1. basic-template.md
   - Simple, beginner-friendly structure
   - Minimal parameters
   - Clear annotations explaining each component
   - Example: "Medium shot of instructor at desk"

2. advanced-template.md
   - Complex, production-ready structure
   - All optional parameters
   - Professional cinematography terms
   - Example: "Medium shot (f/2.8), professional instructor..."

3. examples.md
   - 5-10 real-world examples
   - Platform-specific variations (HeyGen, Veo, Runway)
   - Cost estimates
   - Quality tier recommendations

4. customization-guide.md
   - How to adapt templates for different subjects
   - Variable substitution patterns
   - Platform migration guide

Expected Output: 12 subcategories × 4 files = 48 template files

Agent 7: ai-curriculum-specialist​

Role: Validate educational prompts and ensure pedagogical soundness Tools: Read, Write, Edit, Bash, Grep, Glob, TodoWrite Why This Agent:

  • ✅ Curriculum development expert - Master of educational architecture
  • ✅ Learning objective alignment - Ensures prompts support learning goals
  • ✅ Quality validation - Reviews content for pedagogical effectiveness
  • ✅ Standards compliance - Checks against educational best practices

Specific Task:

Review all educational prompts (categories 01-educational-content):

1. Verify alignment with Mayer's Multimedia Learning Principles
2. Ensure appropriate cognitive load (6-minute video length guideline)
3. Validate accessibility considerations
4. Check for clear learning objectives
5. Assess engagement and retention strategies
6. Recommend improvements for educational effectiveness

Expected Output:

✅ 45 educational prompts reviewed
✅ 3 prompts flagged for cognitive overload (simplified)
✅ 2 prompts enhanced for accessibility (captions guidance added)
✅ All prompts aligned with learning science principles
✅ Quality score: 92/100 (enterprise-grade)

Phase 4: Coordination & Quality Assurance​

Agent 8: orchestrator (Coordinator)​

Role: Coordinate all 7 agents across 4 phases Tools: TodoWrite, Read, Write, Edit, Grep, Glob, Bash Why This Agent:

  • ✅ Multi-agent coordination specialist - Designed for complex workflows
  • ✅ Token budget management - Prevents context collapse
  • ✅ Recursive workflow execution - Handles cascading tasks
  • ✅ Quality gate enforcement - Ensures standards compliance

Specific Task:

Phase 1: Orchestrate Research Extraction
  1. Launch codebase-locator → get file:line references
  2. Launch codebase-pattern-finder → extract prompts (depends on step 1)
  3. Launch codebase-analyzer → analyze patterns (depends on step 2)
  4. Quality gate: Verify 100+ prompts extracted

Phase 2: Orchestrate Content Organization
  1. Launch documentation-librarian → organize prompts
  2. Launch codi-documentation-writer → generate docs (depends on step 1)
  3. Quality gate: Verify 34 documentation files created

Phase 3: Orchestrate Template Generation
  1. Launch educational-content-generator → create templates
  2. Launch ai-curriculum-specialist → validate (depends on step 1)
  3. Quality gate: Verify 48 template files created

Phase 4: Final QA & Delivery
  1. Comprehensive review of all 82 files
  2. Cross-reference validation
  3. Generate master index
  4. Create continuation plan

Expected Output:

✅ Phase 1 Complete: 120+ prompts extracted and analyzed
✅ Phase 2 Complete: 34 documentation files generated
✅ Phase 3 Complete: 48 template files created
✅ Phase 4 Complete: 82 total files, all quality gates passed
✅ Deliverable: Production-ready prompt library

📊 Agent Selection Rationale​

Why NOT Use Other Agents?​

❌ research-agent

  • Too broad: Designed for web search and external research
  • Not optimized: Would search web instead of analyzing local documents
  • Better fit: codebase-pattern-finder for local pattern extraction

❌ web-search-researcher

  • External focus: Searches web sources, not local files
  • Redundant: We already have 151,000 words of research locally
  • Use case: Would only use if researching NEW prompt techniques not in our docs

❌ thoughts-analyzer

  • Wrong domain: Designed for strategic analysis of research notes
  • Overkill: We need prompt extraction, not strategic synthesis
  • Better fit: codebase-analyzer for technical prompt analysis

❌ generative-ui- agents*

  • UI-specific: Designed for React/Vue/Svelte component generation
  • Wrong domain: We're generating prompts for video, not UI code

⚡ Execution Strategy​

Sequential vs Parallel Execution​

Phase 1: SEQUENTIAL (agents depend on each other)

codebase-locator → codebase-pattern-finder → codebase-analyzer
     (find WHERE)        (extract WHAT)           (analyze HOW)

Phase 2: PARALLEL (agents work independently after Phase 1)

documentation-librarian     ← organizes extracted prompts
codi-documentation-writer   ← generates docs (can work simultaneously)

Phase 3: SEQUENTIAL (validation depends on generation)

educational-content-generator → ai-curriculum-specialist
        (create templates)            (validate quality)

🎯 Success Criteria​

Quantitative Metrics​

  • ✅ 120+ prompts extracted from research
  • ✅ 82 total files created (34 docs + 48 templates)
  • ✅ 100% coverage across all 5 categories
  • ✅ Zero duplicates across categories
  • ✅ Quality score 90+ on all documentation

Qualitative Metrics​

  • ✅ Production-ready templates (copy-paste usable)
  • ✅ Clear navigation system (users find prompts in <2 minutes)
  • ✅ Educational soundness (aligned with learning science)
  • ✅ Platform compatibility (HeyGen, Veo, Runway variations)
  • ✅ Integration guidance (works with CODITECT workflows)

💡 Key Insights​

Why This Team Composition is Optimal​

  1. Specialized Skills: Each agent has a distinct, non-overlapping role
  2. Complementary Tools: Different tool access matches different needs
    • codebase-locator: Grep, Glob, LS (search)
    • codebase-pattern-finder: Read, Write, Edit (extract and organize)
    • codebase-analyzer: Deep technical analysis (understand)
    • documentation-librarian: Bash, organization (structure)
    • codi-documentation-writer: Comprehensive writing (document)
    • educational-content-generator: Multi-level creation (teach)
    • ai-curriculum-specialist: Pedagogical validation (quality)
  3. Efficient Workflow: Sequential where dependencies exist, parallel where possible
  4. Quality Assurance: Multiple validation checkpoints throughout process
  5. Avoid Redundancy: No overlapping agent capabilities

Alternative Approaches (Less Optimal)​

❌ Single Agent Approach: Use orchestrator for everything

  • Problem: Token limit exceeded, context collapse, no specialization
  • Better: Divide work among specialists

❌ Too Many Agents: Add research-agent, web-search-researcher, etc.

  • Problem: Redundant work, coordination overhead, diminishing returns
  • Better: Minimal effective team (7 agents is optimal)

❌ Wrong Agent Mix: Use generative-ui agents for prompts

  • Problem: Wrong domain expertise, poor quality output
  • Better: Use domain-specific agents (educational, documentation)

📋 Implementation Checklist​

Pre-Execution​

  • Verify all 9 research documents accessible
  • Confirm directory structure created (17 directories)
  • Allocate token budget (50K per phase = 200K total)
  • Set up quality gates and acceptance criteria

Phase 1 Execution​

  • Launch codebase-locator for file:line discovery
  • Launch codebase-pattern-finder for prompt extraction
  • Launch codebase-analyzer for technical analysis
  • Quality gate: 100+ prompts extracted ✅

Phase 2 Execution​

  • Launch documentation-librarian for organization
  • Launch codi-documentation-writer for docs generation
  • Quality gate: 34 documentation files created ✅

Phase 3 Execution​

  • Launch educational-content-generator for templates
  • Launch ai-curriculum-specialist for validation
  • Quality gate: 48 template files created ✅

Phase 4 Execution​

  • Orchestrator runs final QA review
  • Cross-reference validation complete
  • Master index generated
  • Continuation plan created
  • Quality gate: 82 total files, all standards met ✅

Recommended Execution: Use orchestrator agent to coordinate this entire strategy automatically.

Command:

Task(
  subagent_type="orchestrator",
  prompt="Execute the 4-phase prompt library generation workflow using the 7-agent team defined in agent-team-strategy.md"
)

Expected Outcome: Complete, production-ready prompt library with 82 files in 4-6 hours of agent work.

Agent Team Strategy v1.0 - Optimized for prompt library generation from 151,000+ words of research