• 

  Landing page

• 

  Idea Freewriting Phase

• 

  Idea Refinement Phase

• 

  Deep Research Phase

• 

  Multi-Agent Debate Phase

https://github.com/thorchh/ClaudeUSCHacks

🧠 BrainstormAI

As a solo founder constantly juggling half‑baked ideas and scattered feedback - from Slack threads to Figma comments - I needed a unified co‑pilot. I envisioned an AI that could not only capture my raw thoughts but systematically reason, challenge assumptions, and deliver a clear, actionable plan. Enter BrainstormAI powered by Claude 3.7

🔍 What it does

BrainstormAI is your end‑to‑end ideation pipeline:

Divergent Capture – Free‑text input to spark raw creativity

Context & Feasibility Check – Structured JSON intake of goals, constraints, audience; instant Chain‑of‑Thought (CoT) feasibility scoring (technical, market, financial, regulatory)

Concept Clarification – Parses keywords, objectives, assumptions; auto‑asks follow‑ups for missing detail

Meta‑Creative Expansion – Applies SCAMPER, Six Thinking Hats, First Principles, and phenomenological probing to generate fresh angles

Cross‑Pollination – Fuses your idea with other concepts and analogies (“What if Stripe met Slack?”)

Multi‑Agent Debate – Market, Feature, Synthesis, Contrarian agents present, critique, vote, and refine over Delphi rounds

Report Generation – Outputs structured JSON plus human‑friendly summaries (theme clusters, risk flags, feature matrices, SWOTs, MVP blueprint) ready to export

🛠 How I built it

Anthropic Claude 3.7: Leveraged Claude’s advanced CoT capabilities for transparent, multi‑step reasoning and safe, aligned responses.

JSON‑Schema Tools: Defined each agent (extract_themes, intake_context, meta_creativity, cross_pollination, etc.) with strict input/output schemas for machine‑readable pipelines.

run_claude_tool() wrapper: A simple Python function to sequence Claude tool calls, enforce schemas, and parse outputs.

Chain‑of‑Thought Prompts: Carefully designed templates to get Claude into structured reasoning for deep analysis.

Debate & Voting Logic: Simulated a mini‑Delphi process where Claude agents anonymously critique and re‑vote, ensuring robustness.

⚠️ Challenges I ran into

Prompt Orchestration: Balancing open creativity with strict JSON output - too loose invites drift, too rigid stifles ideation.

State Management: Passing evolving context through seven stages in a single‑file script while preserving immutability.

Latency vs. Depth: Chain‑of‑Thought and multi‑agent debates yield richer insights but can add API latency - tuning max_tokens and agent count was key.

🏆 Accomplishments I’m proud of

Solo MVP: A fully functional prototype that goes from freewriting to exportable roadmap.

Multi‑Agent Creative Stack: Integrated psychological frameworks (SCAMPER, Six Hats) plus analogical synthesis and contrarian checks.

Structured, Exportable Outputs: Clean JSON payloads for UI integration and human‑friendly summaries for user consumption.

Adversarial QA Layer: Automated contrarian agent that surfaces hidden risks and strengthens final recommendations.

📚 What I learned

Modularity Wins: Small, focused agents are easier to debug, extend, and replace.

COT + Schema = Clarity: Claude’s transparent reasoning combined with JSON schemas gives both auditability and machine‑readability.

User‑Centric Prompts: Tailoring the prompt to ask “why” and “what if” unlocked deeper, more actionable insights.

🚀 What’s next for deepBrainstormAI

Real‑Time Integrations: Live connectors for Gmail, Slack, Figma, GitHub to auto‑ingest feedback.

Interactive UI Prototype: Tabbed pipeline with inline “rerun” buttons, confidence gauges, and editable fields.

Learning Loop: Persist past projects to inform Claude’s priors, trend detection, and analogical matches.

Expanded Research Agents: Add live news scraping, social‑sentiment analysis, patent landscape mapping, and startup funding signals.

Built With

• claude
• flask
• next.js
• python