Skills

Arena: Make 100 Claude Agents Fight to Get You a Better Answer

6 minute readUpdated October 2026Explore more

TL;DR

Arena is a free Claude Code skill that stops the re-prompting loop. It gives the same task to 100 Claude sub-agents, each with its own way of thinking, makes them attack each other's answers, and has a judge knock out the loser each round until one answer is left. All 100 agents are the same model, and a full run uses a lot of tokens, so use quick mode (16 agents) for everyday jobs.

Learn Claude Code. Earn income. Only $9.

👉 https://www.skool.com/claudecodeclub

What the Arena skill does: 100 Claude agents, one winning answer

Arena is a free skill for Claude Code made by Jake Schincariol. When Claude gives you an answer you don't like, you type /arena instead of asking again and again. Arena gives the exact same task to 100 sub-agents, runs a knockout tournament between their answers, and brings back the one that survived.

How the Arena tournament works

The Arena skill repo on GitHub with its README
Arena skill by Jake Schincariol on GitHub (MIT license)

What Arena does for you: every agent gets the same task, word for word, plus its own strategy card: a way of reasoning, a workflow and a strategy. So the 100 answers come out different. Then the agents pair off. Each one attacks the other's answer, defends its own and rewrites it. A separate judge agent scores both on a written rubric, and the loser is out. 100 becomes 50, then 25, 13, 7, 4, 2 and 1.

You get back the winning answer, the attacks it beat on the way, and how many rounds it took. If you started from an answer you rejected, a last judge compares the winner with it, and Arena tells you the score even when your old answer wins.

How to install Arena: copy the repo URL below, paste it into Claude Code, and say: install this skill, then confirm /arena works.

repo URLhttps://github.com/Jakeschincariol/arena-skill

Or install it by hand (from the README). Arena needs Claude Code and Python 3.8 or newer, and there is nothing to pip install.

bashgit clone https://github.com/Jakeschincariol/arena-skill.git
cp -r arena-skill/skills/arena ~/.claude/skills/

Best practices for Arena:

  • Use quick mode for everyday jobs. /arena --quick runs 16 agents and about 91 sub-agent calls. A full 100-agent run is about 595 calls, so save it for the answer that matters.
  • Write a clear task first. The agents only see a task file, not your chat. A vague task gives you 100 flavors of vague.
  • Turn on accept-edits mode (Shift+Tab). Every agent writes a file inside .arena/, and the default mode asks you to approve each one.
  • Read the attacks, not just the winner. The winning answer comes with the attacks it survived, which shows you where a weak spot was.

Prompts to try with Arena:

prompt/arena --quick write the headline for our pricing page. We sell [your product] to [your customer]. Keep it under 10 words.
prompt/arena --agents 32 fix the flaky test in [your test file]. Show the fix as a diff, and do not edit my files.

Which Arena setting to start with

  • Everyday writing, ideas, small fixes: /arena --quick (16 agents).
  • A harder problem: /arena --agents 32 (about 187 calls).
  • The one answer that really matters: plain /arena with all 100 agents.
Learn Claude Code. Earn income. Only $9.

👉 https://www.skool.com/claudecodeclub

Common questions

  • Are the 100 agents really 100 different AI models?

    No. They are 100 sub-agents of the same Claude model you are running. What makes them different is the strategy card each one gets: a way of reasoning, a workflow and a strategy.

  • Is the Arena skill free?

    The skill is free, MIT licensed, with no signup and no API key. The tokens are yours, though, and 100 agents use a lot of them.

  • How many agents does quick mode run?

    16. According to the README, that is about 91 sub-agent calls, compared with about 595 for the full 100.

  • Will Arena change my files?

    No. Every agent writes inside a .arena/ folder. The winning answer comes back to you, and using it is your call.

  • Does the winner have to be right?

    No. The judges are Claude too, scoring against a written rubric. You get the strongest answer the tournament found, not proof that it is correct.

Want the skills that make Claude Code worth every token?

Get 650+ plug-and-play skills, MCPs & prompts, plus 8,000+ members - $9/mo, cancel anytime.

Join the Club