Stop Reprompting the Same Claude
When an answer comes back mediocre, the usual move is to tweak the prompt and try again — sometimes for half an hour. arena-skill takes the opposite approach. It's a free Claude Code skill: run /arena once and it sends the same task to as many as 100 sub-agents at the same time, then lets a tournament decide which answer wins.
Same Task, Different Strategy Cards
Every sub-agent gets the exact same task but a different strategy card, so you don't get 100 copies of one answer — you get 100 genuinely different attempts. The cards cover reasoning styles (first-principles, inversion, contrarian), workflows (draft → critique → rewrite, test-first) and tactics (maximal rigor, edge-case hunting), and they combine into thousands of possibilities.
One thing worth knowing: these are all sub-agents of the model you're already using. The variety comes from the strategies, not from different models.
A Judge Picks the Winner
A separate judge sub-agent, which never sees the strategy cards, scores every answer against a strict rubric — correctness, completeness, robustness, specificity and clarity. The weakest answers are eliminated each round until a single one survives.
Mind the Cost
A full 100-agent arena is expensive: the guide puts it at roughly 595 sub-agent calls. For everyday tasks, /arena --quick runs 16 agents instead. Save the full arena for the questions where the best answer really matters.
What's in the Guide
How the strategy cards work, how the judge scores and eliminates answers, the install commands, the quick mode for everyday use, and a direct link to the repo.
The full field guide is in the PDF.
Get the Guide
Drop your email below and we'll send it straight to your inbox.