AI Debate Club — AI skill for Claude Code

AI Debate Club

A multi-agent debate engine. Two different LLMs argue any motion, research the web for themselves, and cite their own evidence. Two independent auditors fact-check them, and a judge decides. Built with LangGraph on Amazon Bedrock, with local HuggingFace models and SQLite memory. Every motion is debated; there is no refusal stage.

Pipeline

1. Intent: the input is spelling-corrected and searched once on Tavily as a single phrase. Claude Haiku 4.5 frames the motion: both sides are kept whole, FOR is the first-named side, and the two positions are direct opposites. Hedged positions and swapped or dropped sides are caught in code, which reruns the framing.
2. Debate (N rounds): Claude Haiku 4.5 and Amazon Nova 2 Lite are research agents. Which one argues FOR is drawn per debate, to avoid a model bias. Each searches the web itself (Tavily tool), keeps its own notebook of F# or A# passages across turns, and cites it. A repetition guard re-asks a side that copies its previous turn.
3. Style: DeBERTa v3 zero-shot (local) scores each turn's emotional style, for display.
4. Fact-check: claims are extracted, and uncited ones are traced back to their source sentence. Sonnet 4.6 and Haiku 4.5 check them in parallel, each claim only against the passages it cites, and a quoted sentence must exist in the passage. Verdicts are reconciled in code.
5. Judge: Claude Sonnet 4.6 decides from the reconciled audit and the transcript.

Memory (SQLite, debate_club.db)

• Session: LangGraph checkpointer, one thread per debate.
• Overall: debates, turns, notebook passages, claims, every step's output, and a Tavily cache, so repeated searches cost no credits. The History box replays any saved debate with no model calls.

Setup

python3.11 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
cp .env.example .env        # add your TAVILY_API_KEY

AWS credentials need Bedrock access to C