When Claude hands you something kind of meh, you usually ask again, and again. The Arena skill does something else: it gives the same task to a crowd of Claude sub agents, each one working it a different way, and makes them attack each other's answers until one is left standing. This guide shows you how to install it, when to use it, and how to keep it from eating your usage.
Checked against the repo github.com/Jakeschincariol/arena-skill (MIT license, made by Jake Schincariol), October 2026. If a step here differs from the repo, the repo is right.
/arena. Behind the scenes it runs a single elimination bracket.| STAGE | WHAT HAPPENS |
|---|---|
| Spawn | Every sub agent gets the exact same task, word for word, plus its own strategy card: a way of reasoning (first principles, inversion, contrarian...), a workflow (outline first, three drafts and pick one...) and a strategy (simplest thing that works, edge cases first...). No two cards are the same. |
| Attack | Answers are paired off. Each side tears into the other's answer: what is wrong, what it missed, what breaks it. |
| Defend | Each side answers every attack it took and rewrites its own answer to fix what it conceded. |
| Judge | A separate judge scores both rewritten answers on a written rubric: correctness, completeness, robustness, specificity and clarity. The loser is out. |
| Repeat | The field halves every round. With 100 agents that is 100, 50, 25, 13, 7, 4, 2, 1. |
| Result | You get the one answer that survived, the attacks it beat on the way and how many rounds it took. |
.arena, and only the winning answer comes back to you.
Open Claude Code and send this message as it is:
Run these two lines inside Claude Code. Installed this way the command is called
/arena-skill:arena.
In a terminal. This gives you the plain /arena command in every project:
.arena. In the default permission mode
that is one approval click per file, which is hundreds on a full run. Press Shift+Tab
until you see accept edits, then start the arena.
/arena right after an answer you did not like, or give it a task. Replace everything in [BRACKETS].| ADD THIS | WHAT IT DOES |
|---|---|
--quick | 16 competitors instead of 100. The everyday setting. |
--agents 32 | Pick the number of competitors yourself. |
--seed 7 | Same seed, same cards and same bracket. Not the same answers. |
| MINUTES | DO THIS | WHY |
|---|---|---|
| 0 to 5 | Install it with the paste the link message and check that /arena answers | Proves it works before you spend anything |
| 5 to 10 | Pick one thing Claude got wrong this week and write the task with every detail | The task decides the result |
| 10 to 40 | Turn on accept edits and run it with --quick | 16 agents, about 91 sub agent calls |
| 40 to 50 | Read the winner and the attacks it survived | You see why it won, not just what it says |
| 50 to 60 | Use it, or tell Claude what is still off and run it again | Your call, always |
--quick (91 calls) for everyday things and save
the full 100 for the answer that really matters. The skill prints the size of the run before it starts.