osolmaz

AI Battle — AI skill for Claude Code

AI community

Structured judged competition between AI agents where they try to pose each other problems the other side should not be able to answer.

How to install AI Battle

This entry records only its repository, not the path inside it, so there is no exact command to give. Open osolmaz/ai-battle and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What AI Battle does

Structured judged competition between AI agents where they try to pose each other problems the other side should not be able to answer.

Alternatives in AI

Debate — AI Debate Hub - Structured four-way debates between Claude, Sonnet, Gemini, and Codex 2.8k ★
Wiki Query — Query the LLM Wiki and synthesize an answer 2.3k ★
Pro Workflow — 1,400+ Battle-tested Claude Code workflows from power users 1.4k ★

README

AI Battle

AI Battle is an experiment in structured competition between AI agents.

Participant agents compete by asking hard questions, answering difficult questions well, and trying to show stronger reasoning than the other side. A judge agent decides who gets the point after every answer.

The competition works like this:

one participant asks a question
the other participant answers it
the judge decides who won that turn
the score is updated immediately
then the roles switch and the next turn begins
by default, each participant asks 10 questions, so a standard match has 20 turns total

Over time, each side has to prove two things: that it can ask clean, challenging questions, and that it can answer difficult questions well. The match is meant to reward both offensive skill and defensive skill, not just one or the other.

Each participant gets up to `30` minutes for each ask turn and each answer turn.

If a participant misses that limit, the match runner sends one final `1`-minute message telling them to return the final JSON immediately. If they still do not return a valid result, they automatically lose that turn and the match continues.

Current Scores

Codex vs Claude

These are the full recorded Codex vs Claude runs so far.

Recorded setup for these runs:

codex participant and codex judge ran through codex-acp and used model gpt-5.4 with reasoning effort xhigh
claude participant ran through claude-agent-acp and used model claude-opus-4-6 with high thinking

Run	Opener	Codex Score	Claude Score	Winner
2026-04-04-0654-codex-vs-claude	`codex`	13	13	`tie`
2026-04-04-1524-codex-vs-claude	`codex`	11	9	`codex`
2026-04-04-1950-claude-vs-codex	`claude`	13	13	`tie`
[2026-04-04-214