Ralph Brainstormer is a multi-agent planning engine that uses a "Debate & Consensus" workflow to generate high-quality technical roadmaps. It pits multiple AI models (Claude, Gemini, Codex) against each other to critique, refine, and ultimately agree on a winning plan.
- Drafting Phase (Collaborative): Generates 6 unique plans (2 iterations x 3 AI models).
- New: Agents are aware of each other's drafts in Round 2. "Claude proposed X, so I will propose Y to complement it."
- Ranking (The Ruthless Judge): All models vote on the drafts using a strict 0-100 scoring system.
- Generic plans are penalized.
- Technical depth is rewarded.
- Synthesis (Cross-Pollination): The Top 3 winners are merged into a single "Unified Blueprint".
- The synthesizer explicitly combines the best ideas from different agents (e.g., "Use Gemini's database schema with Claude's API design").
- Boardroom Critique:
- Agents act as "Senior Architects" to find fatal flaws in the Unified Blueprint.
- Final Polish: The plan is rewritten to address all critiques, resulting in a Master Plan.
- Python 3.8+
- Access to AI CLIs:
- Claude:
claude(Claude Code) - Gemini:
gemini(Gemini CLI) - Codex:
codex(vianpm install -g @shadow/codexor similar) - MUST be installed, simulation is disabled.
- Claude:
- Clone the repository:
git clone https://github.com/alrightryanx/ralph-brainstormer.git cd ralph-brainstormer - Install dependencies:
pip install -r requirements.txt
Run the engine with one click using run.bat (Windows) or via terminal:
python ralph_brainstormer.py --project "MyProject" --objective "Build a decentralized chat app"The final plan will be saved to plans/ralph_brainstorm/MyProject/FINAL_MASTER_PLAN.md.
MIT