Jakeschincariol/arena-skill

The Arena skill – a Claude‑code tournament for better answers

What it does

• The skill lets you run a tournament of many Claude sub‑agents on the same prompt. Each sub‑agent gets the same task but a different “strategy card” (a reasoning mode, workflow and strategy). They then attack, defend, and are judged against each other until only one answer remains.
• The final output is the surviving answer together with the attacks it survived, the card it used, and how many rounds it took.

Why it exists

• Claude can sometimes give unsatisfying replies. By spawning many variants that reason differently, the skill hopes to surface a more robust, complete answer without you having to rewrite the prompt repeatedly.

How you use it

1. Install the skill in Claude Code (copy the arena folder into ~/.claude/skills/ or add it as a plugin).
2. Run the slash command, e.g.:

   /arena --agents 32 fix the flaky test in tests/test_api.py
   

   • --agents N sets the number of competitors (default 100).
   • --quick runs a smaller 16‑agent tournament for everyday use.
   • --seed S makes the card distribution deterministic.
   • --wave W changes how many sub‑agents are launched concurrently.
3. The skill creates a hidden .arena/ directory, writes each sub‑agent’s work to JSON files, and orchestrates the bracket using the pure‑Python bracket.py state machine.
4. When the tournament ends you receive the winning answer and a short report.

Key components

• bracket.py – a standard‑library Python script that tracks the tournament tree, pairs agents, records attacks, and decides winners.
• strategies.json – defines 15 reasoning modes, 12 workflows and 12 strategies (2,160 possible unique cards). The dealer hands out cards without repeats.
• rubric.md – the scoring rubric used by a judge sub‑agent (correctness, completeness, robustness, specificity, clarity).
• Tests – a unit‑test suite runs a full 100‑agent tournament with mock sub‑agents to verify the bracket logic.

Cost

• The skill itself is free and MIT‑licensed. You pay the Claude token usage for every sub‑agent call (e.g., a full 100‑agent run makes 595 calls). Use --quick or fewer agents to keep costs low.

Limitations

• It only works inside Claude Code because it relies on the Agent tool to spawn sub‑agents.
• No parallelism beyond Claude Code’s concurrency limit (default 10 concurrent tool calls). Larger waves require raising CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY.
• The tournament never touches your project files; all intermediate data lives in .arena/.
• The final answer is the most surviving one, not a guarantee of correctness.

Who made it

• Jake Schincariol (opusjake.ai).

License

• MIT – you can modify and redistribute.

---

The Arena skill is a genuine, Claude‑centric AI‑agent orchestration tool that lets you harness many parallel reasoning paths to improve answer quality.