It spawns the 100 sub-agents. The Arena does not run in the normal Claude chat.
// you asked for this
A free Claude Code skill for when Claude gives you a mid answer: 100 sub-agents take the same task, attack each other's work, and only the strongest answer survives. Here is the repo and the setup.
Open it, and see what is inside.
Three cards you copy and paste into any AI chat.
The usual problems, and the fix for each.
// get it
When Claude gives you a mid answer, the Arena gives the exact same task to 100 sub-agents, each with a different strategy card. They pair off, attack each other's answers, defend and rewrite, and a judge scores every match. 100 becomes 50, 25, 13, 7, 4, 2, 1.
Free, MIT licensed, no signup and no API key. It never touches your files.
Open The Arena GuidePDF . opens in a new tab11 pages: what you need, three ways to install, every command, the bracket and what it costs in tokens, how to write a task worth 100 agents, and how to read the winner.
Open The Arena On GitHubgithub.com/avi691/arena-skillIt spawns the 100 sub-agents. The Arena does not run in the normal Claude chat.
Check with python3 --version. Nothing to pip install.
Each agent saves a file. Press Shift + Tab in Claude Code before a run, or approve each one.
// set it up
All three go inside Claude Code. Tap Copy, paste, send.
Step 1 of 3 . start here
Paste this into Claude Code. It installs the skill and checks /arena works.
https://github.com/avi691/arena-skill Install this skill, then confirm /arena works.
Step 2 of 3
16 agents instead of 100. Put the full task after the command.
/arena --quick [YOUR TASK, WITH EVERY REQUIREMENT IN IT]
Step 3 of 3
Right after Claude gives you an answer you do not like, type this. The rejected answer becomes the one to beat.
/arena
// get unstuck
Restart Claude Code after installing. Installed as a plugin, it is called /arena-skill:arena.
Every agent saves a file into .arena/. Press Shift + Tab to switch Claude Code to accept-edits mode before a run.
The skill is free, the tokens are yours. 100 agents is 595 sub-agent calls; --quick (16 agents) is 91. Run bracket.py plan --agents N to price a run first.
The agents only see the task file, not your chat. Put every requirement in the task, or check .arena/<run>/task.md before it starts.
Check the repo's README, or email support@variancemarketing.com.
// what next . paid . optional

You came to get better answers out of Claude. The Playbook shows how to point AI at your whole business: ads, sales and delivery.
I wrote The AI Scale Playbook for the rest: ads, SEO, your website, cold email, DMs, AI calling, video, sales and finance.
All sales are final. You get instant access to digital files, so there are no refunds.
See what is in the Playbook$27 . one payment . next pageNot for you? Close this page. Everything above stays free.