Make 100 Versions of Claude Fight to the Death

When Claude keeps giving you answers that miss, asking again usually gets you another version of the same miss. The Arena skill hands one task to up to 100 Claude sub-agents, has each one attack it a different way, then makes them fight until one answer is left standing.

Start Here: How the Arena Works

If you came from the ARENA video, this is the full setup. Install it once, then reach for it whenever Claude's answer is not where you want it to be.

You type one command. Claude spins up sub-agents (100 by default) and gives every one of them the exact same task, word for word. Each one also gets its own strategy card: a way of reasoning, a workflow and a strategy. The card is what makes each agent come at the problem from a different angle.

Then they pair off. Each agent attacks its opponent's answer, defends its own, and rewrites it to fix whatever it conceded. A separate judge scores the match on a written rubric and the loser is out. 100 becomes 50, then 25, 13, 7, 4, 2 and finally 1.

I tested it by giving 100 of them one job: write the perfect hook for my next video. The hook that survived beat everything I had come up with in the previous 24 hours of writing.

Inside the Arena: Cards and Judging

Two things decide how a run plays out: the card each agent is dealt and the rubric the judge scores with.

What's on a strategy card

That adds up to 2,160 possible cards, and no two agents get the same one. All 100 agents are the same Claude model you are already using. The cards are the only thing that makes them different.

How the judge scores each match

An answer with a verified fatal flaw cannot beat one without, whatever the totals say. The judge never sees the strategy cards, and a longer answer earns nothing for being longer.

Copy This First: The One-Paste Install

Open Claude Code and paste this. It is the install message from the skill's own instructions.

https://github.com/Jakeschincariol/arena-skill

Install this skill, then confirm /arena works.

Claude should download the skill and copy its folder into your personal skills folder, then check that it loads. Approve the steps it asks about. If /arena does not appear when you type a slash, start a fresh Claude Code session and check again.

Prefer to install it yourself?

Run these commands in your terminal. They download the repository and copy the arena folder into your personal skills folder, so the skill works in every project. The first line just makes sure that folder exists.

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

Or install it as a plugin

Run these two commands inside Claude Code.

/plugin marketplace add Jakeschincariol/arena-skill
/plugin install arena-skill@arena-skill

As a plugin, the command shows up as /arena-skill:arena instead of /arena. Use the folder copy above if you want the short version.

Type / in Claude Code and look for arena in the list. Then ask Claude to print the plan for a run. The plan command only does the math. It does not start an arena or write any files.

Find the Arena skill's bracket.py and run its plan command twice: once with --quick and once with --agents 100. Show me both tables. Do not start an arena.

You should see two tables. The quick one totals 91 sub-agent calls over 4 rounds. The full one totals 595 calls over 7 rounds.

If you copied the folder yourself, you can run the plan straight from your terminal:

python3 ~/.claude/skills/arena/bracket.py plan --agents 100

If that says Python is missing or too old, install a current version of Python 3 first. There is nothing else to install.

Every agent saves its work as a file inside .arena. In Claude Code's default permission mode, that means one approval per file, which adds up to hundreds of prompts on a big run.

Before you start, press Shift+Tab until Claude Code shows that it will accept edits. The skill reminds you about this, but it will not change your settings for you.

If the task involves code, run the arena from that project so the agents can read your files. They still write only inside .arena, and code changes come back to you as a diff or full files for you to approve.

There are two ways to start one.

After a bad answer. Just type /arena. It takes your last request as the task and the answer you did not like as the one to beat.

With a fresh task. Type /arena, any options, then the task. Start with --quick, which runs 16 agents over 4 rounds.

Try it on a hook

/arena --quick Write one opening hook for my next short-form video. The video is about [topic]. My audience is [who they are and what they care about]. The hook has to work spoken out loud in about [number] words and make a specific promise the video pays off. Give the final hook plus two backup versions.

Claude prints the size before it starts, such as "16 agents, 4 rounds, 91 sub-agent calls." After each round you get one line, such as "Round 2 done: 4 of 16 left." The attacks and verdicts stay out of your chat.

Claude may also offer an arena when you say something like "that's a bad answer, make them compete." When it starts itself that way, it asks first and gives you three choices: the full arena, the quick arena or an ordinary retry.

This step decides the result. The agents cannot see your chat. The skill writes a standalone task file at .arena//task.md, and that file is all they ever know. If a requirement never made it into the file, all 100 agents miss it.

Put everything in the task: who it is for, the format, the length, the tone and what "done" looks like. If you rejected an earlier answer, say what you disliked about it in your own words.

/arena --quick [What you want, in one sentence.]

Who it is for: [audience, and what they already know]
Format and length: [for example, one paragraph under 120 words, or a 5-row table]
Must include: [facts, names, numbers, links or files it has to use]
Must avoid: [words, claims, styles or approaches you do not want]
Tone: [for example, direct and warm, like talking to a friend]
Done looks like: [how you will judge whether it worked]
What I disliked about the last answer: [in your own words, or delete this line]

Check the brief before a big run

Before you spend a full 100-agent run, have Claude write the brief in a normal message first. You can read it, fix it, then start the arena.

I want to run the Arena skill on this, but do not start it yet: [describe what you want].

Write a standalone task brief that a stranger could follow without seeing this conversation. Include every requirement and constraint I have given you anywhere in this chat: audience, format, length, tone, what must not change and what done looks like. Add any context they would need, such as file paths or pasted data. Do not add requirements I never gave, and do not include your own view of the right answer. Show me the brief and wait.

Once it looks right, type /arena --agents 100 and paste the approved brief after it. Leaving your own view of the right answer out of the brief matters. Otherwise you push all 100 agents the same way, which defeats the point.

When the bracket finishes, you get:

1. The winning answer, in full.

2. Why it won: the attacks it survived, plus its strategy card on one line.

3. The round count, such as "7 rounds, 100 agents in, 1 left."

4. If you started from a rejected answer, a blind final judge's scores for the winner and your old answer. If the old one scored higher, the skill tells you plainly.

5. Where the full record lives.

If the winner would change files in your project, the skill does not apply it. It asks whether you want to apply it or change it.

Look at the final match

Open the latest run folder in .arena. Show me the final match: both finalists' revised answers and the judge's scores and reason. Do not start a new arena or change any files.

Pick the Right Size

Every call reads the task and one or two answers, so a long task costs more per call. A rejected answer to beat adds one more call at the end. The skill prints these numbers before it starts.

The options

The waves exist because Claude Code runs at most 10 tool calls at once by default, set by CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY. That is why a 100-agent arena runs as 70 waves and takes a while.

What to Throw Into the Arena

The arena works best on a task with a clear goal and a way to judge the result. These are good starting points.

Make a hard decision

/arena --agents 32 Recommend which of these options I should choose: [option A], [option B], [option C]. My goal is [goal]. My constraints are [budget, time, skills]. Give one recommendation, the main trade-off it accepts, and the first actions to take this week.

When to skip it

Skip the arena for quick facts, simple questions and anything one normal retry with better instructions would fix. It also cannot rescue a vague task. Vague task in, 100 flavors of vague out.

Common Mistakes and Easy Fixes

Ask Claude to diagnose it

My Arena skill run is not working. Here is what happened: [describe it or paste the error].

Check where the skill is installed, whether python3 is available and at least version 3.8, and whether an .arena folder exists here. If a run was interrupted, run the skill's status command and tell me where it stopped. Explain the problem in plain words and give me the fix. Do not start a new arena or delete anything without asking me.

Quick Reference: The Arena Skill

Setup in four lines

• Install: Paste the GitHub link into Claude Code with "Install this skill, then confirm /arena works."
• Check: Type / and look for arena, then run the plan command.
• Permissions: Press Shift+Tab to accept edits before a run.
• Record: Everything lands in .arena inside the folder you ran it from.

Give it one task. Let the best answer win.