/arena

name: arena
description: >-
  Make 100 versions of Claude fight to the death over one task. Spins up N
  sub-agents (default 100, --quick for 16), gives every one the exact same
  task plus a different strategy card (reasoning mode, workflow, strategy),
  then runs a single-elimination bracket: they attack each other's solutions,
  defend and revise, and a judge scores every match on a written rubric until
  one solution survives. Use when the user is not satisfied with an answer,
  calls it a bad answer, says try again or do better, says "arena", or asks
  to make them compete.
argument-hint: "[--agents N | --quick] [--seed S] <task>"

arena

For when Claude keeps giving a bad answer. Instead of asking again and again, this runs a tournament: N sub-agents get the exact same task, each attacks it with a different reasoning mode, workflow and strategy, and then they attack each other in a bracket until one solution is left. You are the orchestrator. You never compete and you never judge.

What the user typed after `/arena`: `$ARGUMENTS`

If that is blank, or still reads like a placeholder, nothing was passed: take the task from the conversation.

The tool

Every piece of bookkeeping goes through `bracket.py` in this skill's folder:

python3 "${CLAUDE_SKILL_DIR}/bracket.py" <command>

Below, `ARENA` means exactly that command. If the path looks unexpanded, use the "Base directory for this skill" that Claude Code printed at the top of this skill. The state lives in `.arena/<run>/arena.json` in the current directory, and every command after `init` finds it through `.arena/LATEST`.

Step 1: size it, and get a yes if nobody asked for it

Read the flags out of the request. Everything that is not a flag is the task.

| flag | meaning | | --- | --- | | `--agents N` | N competitors. Default 100. | | `--quick` | 16 competitors. The everyday setting. | | `--seed S` | Fixes the cards and the pairings. Default: random, and recorded. | | `--wave W` | Sub-agents per wave. Default 10. Only raise it if the user raised Claude Code's limit. |

No task text at all means: the task is the user's most recent request in this conversation, and your last answer to it is the baseline to beat.

Run `ARENA plan --agents N` (or `--quick`). It prints the rounds, the sub-agent calls and the waves.

• **The user asked for the arena** (typed `/arena`, said "arena", or asked to make them compete):

tell them in one line how big it is, for example "100 agents, 7 rounds, 595 sub-agent calls", and start.

• **This skill fired because the user is unhappy** ("that's wrong", "try again", "bad answer") and

never mentioned the arena: ask once before spending anything. Offer three options: the full arena (100 agents, 595 sub-agent calls), `--quick` (16 agents, 91 calls), or an ordinary retry. Wait for the answer.

Sub-agents write their work into `.arena/` in the current directory. In the default permission mode that is one approval per file, which is hundreds on a big run. Before the spawn phase, suggest accept-edits mode (Shift+Tab) for the run. Do not change the user's settings yourself.

Step 2: write the task file

This is the step that decides the result. **Sub-agents cannot see this conversation.** Every competitor, attacker and judge knows only what is in the task file, so write `.arena/task.md` to stand on its own:

• The request, in the user's own words where you can.
• Every requirement and constraint the user stated anywhere in the conversation: audience, length,

format, tone, stack, deadline, what must not change.

• The context a stranger would need: absolute paths of the files that matter, pasted data, what the

product is, the conventions in the codebase.

• What "done" looks like, if the user said.
• If there is an answer to beat: what the user disliked about it, in their words.

Do not add requirements the user never gave. Do not write your own view of the right answer into it: that pushes 100 agents the same way, which is the opposite of the point.

If there is an earlier answer the user was not satisfied with, write it word for word to `.arena/baseline.md`.

Step 3: init

ARENA init --agents N --seed S --task-file .arena/task.md --baseline-file .arena/baseline.md

Leave out `--baseline-file` when there is nothing to beat, and `--seed` to get a random one. `init` copies the task into the run folder, deals every competitor a different strategy card with no repeats, pairs round 1, and writes `arena.json`.

Step 4: the loop

Always drive it with `ARENA next`. It reads the state on disk and tells you the next step and the exact command.

Every phase that runs sub-agents works the same way:

1. `ARENA prompts <phase>` writes one brief per job and lists the jobs still to run, in waves. 2. Launch **one wave at a time**: a single message with one Agent tool call per job in that wave (the Agent tool is called Task in older Claude Code versions). Each call is:

• `subagent_type`: `general-purpose`
• `description`: `arena <job id>`
• `prompt`: `Read <prompt path> and follow it exactly. It is your whole brief.`
• `run_in_background`: false, where the tool has that option, so the wave comes back together.

Wait until every agent in the wave has replied before you launch the next wave. 3. After the last wave, run `ARENA next`. If an output is missing it sends you back to the same phase, and `prompts` then lists only the missing jobs. Re-run those once. If a job fails twice, write the single line `NO OUTPUT` into each of its output files (`ARENA check <phase>` lists them) and move on. A missing attack counts as no attacks. A missing solution loses its match. A judge that fails twice gets a third, fresh run: never decide a match yourself.

The order `next` takes you through:

• **spawn**, once: every competitor writes its own solution to the task.
• then every round: **attack** (two per match) → **defend** (two per match) → **judge** (one per