Skip to content

Arena

by Jake SchincariolSkillGet organised

Runs many Claude sub-agents on the same task and keeps the answer that survives a bracket

View on GitHub

The problem it solves

You've asked Claude for the same headline four times and every version still feels flat.

For Claude Code usersPeople with one answer that matters

Before
You ask again and again, and each retry comes at the problem from the same angle as the last.
After
Many sub-agents with different strategies attack and revise each other's answers, and one survivor comes back with the attacks it beat.
Stars
★ 1.1k
License
MIT
Cost
Usage-based
Level
Intermediate
Contents
Includes scripts

Overview

Arena is a Claude Code skill by Jake Schincariol for when you are unhappy with an answer. It writes your task to a file and gives it word for word to 100 sub-agents (16 with --quick), each with a different reasoning strategy. They pair off to attack, defend and revise each other's work while a judge scores every match on a written rubric. You get the one solution left, the attacks it survived and, if you started from a rejected answer, a blind comparison with it.

Instead of asking again, it makes competing answers critique each other, and it tells you plainly if your original answer scored higher.

Try these first

Start with the first one: it is a small task you can check easily.

  1. /arena --quick write a headline for [your page]First try · 91 model calls

  2. /arena --agents 32 plan my launch week, I have [hours] a day187 model calls

  3. That's a bad answer. Make them compete.asks before running

  4. How big would a 100-agent arena be? Show the plan without starting it.read-only

Example prompts written by Agent Flows Hub.

What you get

From the creator's documentation.

OutputWhat it isReads
Strategy cardsEach sub-agent gets its own mix of 15 reasoning modes, 12 workflows and 12 strategiesYour Claude plan
Attack and defendPaired sub-agents attack each other's answers, then concede or rebut and reviseYour Claude plan
Judged bracketA judge scores each match on a five-part rubric you can read and editYour Claude plan
Blind final checkCompares the winner with the answer you rejected, without knowing which is whichYour Claude plan
Winner reportThe surviving answer, the attacks it survived, its card and the number of roundsYour computer
Hands-off filesWorks only inside an .arena folder; changes to your project come back for you to applyYour computer

Good to know

Get a better version of an answer you didn't like, such as a headline or a launch plan. The skill turns the request into a standalone task file, runs a bracket of sub-agents on it and returns the surviving answer with the attacks it beat.

Limits

  • A full 100-agent run makes 595 sub-agent calls in waves of 10, so it takes a while and uses many tokens; --quick runs 16 agents and 91 calls.
  • Every sub-agent writes a file in an .arena folder, so the default permission mode asks once per file. The README suggests accept-edits mode for the run.

Notes by (original project name: arena), written from the creator's documentation, not from running it ourselves. arena SKILL.md · The Arena skill README

Stuck setting it up?

Tell us what happened. We take a look first; if it needs paid work, you get an estimate before anything starts.

Get help

Other skills

  • Article Cover Image Makes a header image for your article in a palette and drawing style you choose
  • Article to Image Slides Turns an article, report or notes into a deck of full-page slide images plus speaker notes
  • Article to Slide Deck Turns an article into a set of slide images, merged into PowerPoint and PDF files
  • Article Translator Translates articles, files or web pages into natural Chinese, English or another language
  • Avoid AI Writing Finds the habits that make English text sound machine-written and rewrites them
  • Banner Designer Designs banners for social media, ads, website headers and print in several styles

See all 140 skills · A skill is installed but not loading

GitHub