GitHub radar
Didn't like Claude's answer? Run 100 copies against each other
This is a skill (add-on) for Claude Code, Anthropic's coding tool. It runs a tournament: 100 copies of Claude solve your task and knock each other out round by round until one proven answer survives.
/arena spins up a hundred Claude copies with different thinking strategies, has their solutions attack and defend each other, and a Claude judge keeps only the best one.
When it helps
Use it when a specific Claude answer let you down and you want something sharper; it won't help with a vague task — it just returns a hundred vague answers.
Pros
- Free, no sign-up or API keys — just Python required
- Never touches your project files — everything goes into .arena/
- Shows you the winning answer plus the attacks it survived
- The --quick mode (16 copies) is cheaper for everyday use
Cons
- Full 100-copy run costs around 600 Claude calls — token-heavy
- It's one AI run a hundred times, not a hundred different models
- Doesn't fix a vague task — just answers vaguely a hundred times
How to set it up — step by step
- 1Open Claude Code inside any of your project folders.
- 2Paste the link https://github.com/Jakeschincariol/arena-skill into the chat and ask Claude to install the skill, then confirm /arena works.
- 3When a Claude answer disappoints you on a real task, run /arena --quick for a cheaper 16-copy tournament instead of the default 100.
- 4Allow edits (Shift+Tab) so Claude Code doesn't ask for approval on every file written to .arena/.
- 5Wait for the final round and review the winning answer along with the attacks it survived.
- 6Decide for yourself whether to apply the suggested answer — the skill never changes your project without your say.
▌ More finds