arena

1# arena
2 
3For when Claude keeps giving a bad answer. Instead of asking again and again, this runs a tournament:
4N sub-agents get the exact same task, each attacks it with a different reasoning mode, workflow and
5strategy, and then they attack each other in a bracket until one solution is left. You are the
6orchestrator. You never compete and you never judge.
7 
8What the user typed after `/arena`: `$ARGUMENTS`
9 
10If that is blank, or still reads like a placeholder, nothing was passed: take the task from the
11conversation.
12 
13## The tool
14 
15Every piece of bookkeeping goes through `bracket.py` in this skill's folder:
16 
17```bash
18python3 "${CLAUDE_SKILL_DIR}/bracket.py" <command>
19```
20 
21Below, `ARENA` means exactly that command. If the path looks unexpanded, use the "Base directory for
22this skill" that Claude Code printed at the top of this skill. The state lives in
23`.arena/<run>/arena.json` in the current directory, and every command after `init` finds it through
24`.arena/LATEST`.
25 
26## Step 1: size it, and get a yes if nobody asked for it
27 
28Read the flags out of the request. Everything that is not a flag is the task.
29 
30| flag | meaning |
31| --- | --- |
32| `--agents N` | N competitors. Default 100. |
33| `--quick` | 16 competitors. The everyday setting. |
34| `--seed S` | Fixes the cards and the pairings. Default: random, and recorded. |
35| `--wave W` | Sub-agents per wave. Default 10. Only raise it if the user raised Claude Code's limit. |
36…