Model Chat

Model Chat

Spawn 5 Claude Sonnet instances into a shared conversation room. They debate a problem across 5 rounds using round-robin turns (all agents respond in parallel each round, see full history). A synthesizer agent then merges the debate into a final answer.

Why this works: Same model, slight framing variations = systematically different failure modes. Surfacing disagreements between instances is more valuable than any single instance's confident answer. Consensus across independent runs filters hallucinations; divergences reveal genuine judgment calls.

Execution

1. Parse the request

Extract from the user's message:

• Topic/problem to debate
• Mode: auto (default) or interactive (--interactive flag)
• Agent count: default 5, user can override
• Round count: default 5, user can override

2. Run the orchestration script

cd /Users/nicksaraev/Business && .venv/bin/python .claude/skills/model-chat/model_chat.py "<topic>"

For interactive mode:

cd /Users/nicksaraev/Business && .venv/bin/python .claude/skills/model-chat/model_chat.py "<topic>" --interactive

Optional flags:

• --agents N — number of agents (default 5)
• --rounds N — number of debate rounds (default 5)

3. Deliver results

The script outputs the full conversation to stdout in real-time and saves to a timestamped subdirectory:

• active/model-chat/<YYYYMMDD-HHMMSS>/conversation.json — full structured conversation log
• active/model-chat/<YYYYMMDD-HHMMSS>/synthesis.md — final synthesized answer
• active/model-chat/latest — symlink to the most recent run

Present to the user:

• Brief summary of key agreements and disagreements
• The synthesis (or link to it)
• Note any particularly interesting moments of debate

How it works internally

1. 5 agents with slight framing variations (systems thinker, pragmatist, edge-case finder, UX-focused, contrarian)
2. Round-robin: each round, all agents see full history and respond in parallel via async API calls
3. 5 rounds of debate (25 total messages)
4. Synthesizer: a final agent reads the entire debate and produces a structured merged answer
5. Interactive mode: after each round, prompts user for optional input to steer the debate

Output files

Each run saves to active/model-chat/<YYYYMMDD-HHMMSS>/:

A latest symlink always points to the most recent run. Previous runs are preserved.