#421421st of 455 in Skills, plugins and prompts
arena-skill
Claude Code skill that runs a sub-agent tournament to pick one answer
- Stars
- 1.5k
- License
- MIT
- Last commit
- Sep 2026
- Language
- Python
121 on 30 Sept 20261,454 on 11 Nov 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
Adds an /arena command to Claude Code that gives one task to N sub-agents (default 100), each with a different strategy card. They attack and revise each other's solutions in pairs, a judge scores each match on a written rubric, and the winner is returned. Bracket state lives in one JSON file managed by a standard-library Python script, and agents write only inside .arena/.
Who it is for: Claude Code users who want a stronger answer after a bad one
Strengths
- Bracket state in one JSON file; resumes via `bracket.py next` after context compaction
- Same task text sent to every agent; 2,160 strategy cards, editable in strategies.json
- Writes only inside .arena/; never edits your project files
- Prints call counts before running; --quick uses 16 agents (91 calls)
Weaknesses
- Default 100-agent run is 595 sub-agent calls, a large token bill
- Judges are the same Claude model, so the winner is not a proof of correctness
- Runs in waves of 10 by default; the full run takes 70 waves
- Needs Claude Code; default permission mode prompts once per written file
What it needs
- no GPU
- Needs Claude Code, Python 3.8+
- Models: Claude (the model running Claude Code)
Also in Skills, plugins and prompts
See all 455| Rank | Project | Score |
|---|---|---|
| 1 | claude-memPersistent cross-session memory plugin for Claude Code and other coding agents | 97 out of 100 |
| 2 | impeccableDesign skill with 24 slash commands for AI coding agents' frontend output | 97 out of 100 |
| 3 | gstackMarkdown slash-command skills that give Claude Code planning, review, QA and ship roles | 96 out of 100 |
| #4 | rufloAgent harness adding swarms, memory and hooks to Claude Code and Codex | 96 out of 100 |
| #5 | cavemanShortens coding-agent replies and compresses tool output before the model reads it | 95 out of 100 |