Agents Crossfire — AI skill for Claude Code

agents-crossfire

Make your AI models argue with each other so you get a better answer.

Instead of asking one model and hoping it's right, crossfire puts Claude, Codex, Gemini, and Ollama in the same room with conflicting personas. They research independently, read each other's arguments, and clash until positions stop moving. You get the debate transcript and a verdict.

Install

npx skills add sderosiaux/agents-crossfire

Or manually: copy `SKILL.md` to `~/.claude/commands/agents/debate.md`

Use

/agents:crossfire should we fine-tune or just use RAG for our domain? ask codex and gemini
/agents:crossfire are AI coding agents ready to replace junior devs? use codex, gemini and sonnet
/agents:crossfire MCP vs tool-use for agent integrations, ask ollama and codex

No flags. Claude parses the topic and participants from your sentence. If you don't name anyone, it defaults to Claude + Codex.

What happens

1. Claude creates a persona for each model (a cautious ops engineer, an aggressive strategist, a skeptical architect) designed to conflict
2. All models research and stake a position in parallel, each writing its own file
3. Each model reads the others' files and reacts directly, minimum 3 rounds
4. Claude writes a verdict with a coalition map showing who aligned with whom

Every response is written to disk by the model itself. Unfiltered, unedited. The debate folder is your audit trail.

What the output looks like

From a "fine-tune vs RAG" debate (Codex vs Gemini vs Sonnet):

COALITION MAP:
Codex + Sonnet aligned: "RAG first, fine-tune only when retrieval
hits a wall." Gemini isolated: "Fine-tuning is underrated because
teams stop at naive RAG and never push past the retrieval ceiling."

KEY EXCHANGE:
Sonnet found that OpenAI's own fine-tuning docs recommend RAG as
the first approach. Codex verified: fine-tuning GPT-4 c