Universal Ideation v3.2

Science-grounded autonomous ideation system with self-improving mechanisms.

Features

• Triple Generator - Explorer, Refiner, Contrarian modes
• 8-Dimension Scoring - Novelty, Feasibility, Market, Complexity, Scenario, Contrarian, Surprise, Cross-Domain
• DARLING Learning - Diversity-aware reward calculation
• Atomic Novelty (NovAScore) - 0.94 accuracy claim-level novelty detection
• Verification Gates - Quality checkpoints
• Reflection Learning - Self-improving pattern extraction
• Plateau Escape - Avoid local optima
• Web Search Integration - Real-time market intelligence via Perplexity API
• Interview Context Integration - Enriched ideation using universal-interview skill
• Batch Mode - Generate multiple ideas per API call (5x faster)

Architecture

Installation

As Claude Code Skill

1. Copy folder to ~/.claude/skills/universal-ideation-v3/
2. Install dependencies:

   pip install -r requirements.txt

3. Invoke with:

   /universal-ideation-v3 "your domain"

Optional: Vector Search

For semantic similarity features:

pip install qdrant-client sentence-transformers
docker run -p 6333:6333 qdrant/qdrant

Quick Start

Via Skill Command

/universal-ideation-v3 "e-commerce innovation"
/universal-ideation-v3 "sustainable opportunity in Malaysia"
/universal-ideation-v3 "ai start up innovation"
/universal-ideation-v3 "sustainable packaging innovation"

Via CLI (Stub Mode)

cd ~/.claude/skills/universal-ideation-v3
python3 scripts/run_v3.py "your domain" --verbose

Via LLM Runner (Full Mode with Claude API)

Requires Anthropic API key in ~/.env:

CLAUDE_API_KEY=sk-ant-xxxxx

Optional: Add Perplexity API key for web search:

PERPLEXITY_API_KEY=pplx-xxxxx

Run with full LLM integration + storage:

cd ~/.claude/skills/universal-ideation-v3
python3 scripts/llm_runner.py "your domain" -i 30 -m 30 -v

This mode:

• Uses Claude API for idea generation and scoring
• Stores ideas in SQLite database
• Stores embeddings in Qdrant for semantic search
• Exports full v3.2 statistics (DARLING learnings, atomic novelty, etc.)

With Web Search (Market Intelligence)

Enable real-time market context from Perplexity:

python3 scripts/llm_runner.py "protein beverages" -i 10 -m 5 -w -v

This fetches:

• Domain trends - Current market developments
• Market gaps - Underserved opportunities
• Emerging tech - Relevant technologies
• Consumer insights - Behavior patterns

The market context is injected into idea generation prompts for more grounded, trend-aware ideas.

With Interview Context (Deep Domain Understanding)

Enrich ideation with structured interview context from the universal-interview skill:

# First, run an interview to build context
cd ~/.claude/skills/universal-interview
python3 scripts/interview_runner.py "sustainable protein beverages"

# Then use the context in ideation
cd ~/.claude/skills/universal-ideation-v3
python3 scripts/llm_runner.py "protein beverages" --context-id [ID] -b -n 10 -v

# Or use interactive context selection
python3 scripts/llm_runner.py "protein beverages" -c -b -n 10 -v

The interview context provides:

• Problem Space - Pain points and user needs
• Constraints - Budget, timeline, regulations
• Assumptions - Hidden beliefs to challenge
• Intent - Strategic goals
• Preferences - What excites vs. bores
• Existing Solutions - Competitors and gaps
• Resources - Assets and capabilities

Both skills share the database at ~/.claude/data/ideation.db.

Batch Mode (5x Faster)

Generate multiple ideas per API call:

# Generate 100 ideas in batch mode
python3 scripts/llm_runner.py "your domain" -b -n 100 -s 10 -v

# With web search
python3 scripts/llm_runner.py "your domain" -b -n 100 -w -v

# With interview context
python3 scripts/llm_runner.py "your domain" -b -n 100 --context-id [ID] -v

Mode	Ideas/min	100 ideas
Standard	~1	~100 min
Batch	~5	~20 min

Options

llm_runner.py (Full Mode)

Flag	Default	Description
-i, --iterations	15	Max iterations (standard mode)
-m, --minutes	15	Max duration (standard mode)
-t, --threshold	60.0	Acceptance score
-v, --verbose	false	Show progress
-w, --web-search	false	Enable Perplexity web search
-b, --batch	false	Enable batch mode
-n, --target	100	Target ideas (batch mode)
-s, --batch-size	10	Ideas per API call (batch mode)
-c, --context	false	Interactive interview context selection
--context-id	-	Specific interview context ID

run_v3.py (v3.3 - LLM Mode)

Flag	Default	Description
-i, --iterations	30	Max iterations
-m, --minutes	30	Max duration
-t, --threshold	65.0	Acceptance score
-v, --verbose	false	Show progress
--use-llm	false	Enable Anthropic API for real idea generation
--llm-model	claude-sonnet-4-20250514	Model to use for generation
--initiative	-	Load enriched_domain from interview database by ID
--test	false	Stub mode for testing

With Interview Integration (v3.3)

Load context directly from interview database:

# Run ideation with interview context
python3 scripts/run_v3.py --initiative e87f16de-61d0-427c-aa21-2cc7b6d3274d --use-llm -v

# Or with manual domain
python3 scripts/run_v3.py "your domain" --use-llm -v

The --initiative flag queries ~/.claude/data/ideation.db for the enriched_domain generated by universal-interview.

Storage

SQLite Database

• Location: ~/.claude/data/ideation.db (shared with universal-interview)
• Tables: ideas, sessions, learnings, initiatives, interview_responses
• Persists all accepted ideas with scores

Qdrant Vector Database (Optional)

• Location: localhost:6333
• Collection: universal_ideas (384-dim embeddings)
• Enables semantic similarity search

Start Qdrant:

docker run -d -p 6333:6333 qdrant/qdrant

Output

Results saved to:

• output/ideation_YYYYMMDD_HHMMSS.json - Full session export
• data/ideation.db - SQLite persistence
• Qdrant vectors - Semantic embeddings (if enabled)

Backup

Backup your ideas database (SQLite + Qdrant vectors):

cd ~/.claude/skills/universal-ideation-v3

# Full backup (SQLite + Qdrant + JSON)
python3 scripts/backup.py backup

# Named backup
python3 scripts/backup.py backup -n "my_backup"

# SQLite only (skip Qdrant)
python3 scripts/backup.py backup --no-qdrant

# View statistics
python3 scripts/backup.py stats

# List all backups
python3 scripts/backup.py list

# Export all ideas to JSON
python3 scripts/backup.py export -o my_ideas.json

# Restore full backup (SQLite + Qdrant)
python3 scripts/backup.py restore backup_file.db

# Restore SQLite only
python3 scripts/backup.py restore backup_file.db --no-qdrant

Backup Files

File	Contents
*_backup.db	SQLite database (ideas, sessions, learnings)
*_qdrant.snapshot	Qdrant vector embeddings
*_backup.json	Full JSON export

Backups saved to backups/ folder.

Testing

python -m pytest tests/ -v

Structure

universal-ideation-v3/
├── SKILL.md              # Skill definition
├── README.md             # This file
├── requirements.txt      # Dependencies
├── scripts/
│   ├── run_v3.py        # Main orchestrator (stub mode)
│   ├── llm_runner.py    # LLM-integrated runner (full mode)
│   ├── backup.py        # Database backup tool
│   ├── generators/      # Triple Generator
│   ├── gates/           # Quality gates
│   ├── evaluators/      # Cognitive diversity
│   ├── learning/        # DARLING + reflection
│   ├── escape/          # Plateau escape
│   ├── novelty/         # Atomic novelty
│   ├── search/          # Web search (Perplexity)
│   └── storage/         # Persistence (SQLite + Qdrant)
├── tests/               # 74 unit tests
├── data/                # Runtime SQLite
├── backups/             # Database backups
└── output/              # Generated results

License

MIT

Version

v3.3 (2026-01-01) - LLM integration for run_v3.py, --initiative flag for interview database integration