arena

Spawns parallel candidate solutions, judges them against a rubric, and synthesizes the strongest result.

Updated Mar 27, 2026
One-click install
npx skills add https://github.com/gmackie/agent-skills --skill arena-gmackie
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: arena
Source: https://github.com/gmackie/agent-skills/tree/main/skills/arena
Command: npx skills add https://github.com/gmackie/agent-skills --skill arena-gmackie

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? A single attempt at a non-trivial artifact often locks in the wrong design shape. This Skill runs multiple independent attempts at the same task in parallel, then merges the best ideas into one verified result instead of betting on one shot. ## Core Features & Use Cases - Parallel Fan-Out: Spawns N subagent candidates on different models, each writing to its own isolated output path with a mandatory rationale. - Rubric-Based Judging: Derives 3-6 concrete gradeable criteria, runs a cross-judge subagent from a different model family, and scores candidates criterion by criterion. - Graft and Verify: Picks the strongest base, folds in the best parts of losing candidates by hand, records rejections, and verifies the synthesized artifact. - Use Case: When designing a non-trivial CLI tool, run the arena to have four models each produce an implementation, then ship a synthesis combining the cleanest base with the best flag-handling approach from a runner-up. ## Quick Start Ask the agent to throw this task in the arena and synthesize the best result from parallel candidates.

Frequently Asked Questions about arena

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I run multiple AI models on the same task and pick the best result?▼

Use the arena workflow: frame a shared prompt with a concrete rubric, spawn N background subagents on different models with isolated output paths, cross-judge the candidates, pick a base, and graft the best ideas from the losers into it.

What is the arena skill used for in Cursor?▼

The arena skill fans out parallel candidate attempts at a non-trivial artifact, judges them against a rubric, and synthesizes one verified result. Trigger it with /arena, 'arena this', or 'throw it in the arena'.

When should I not use parallel candidate generation?▼

Avoid it for trivial or well-specified tasks where one attempt suffices, since the arena adds judging and grafting overhead. If candidates wildly diverge, the task framing was under-specified and should be reframed rather than averaged.

Why does each arena candidate need to write a rationale?▼

The rationale lets the parent tell whether a candidate's structure is principled or accidental, which makes grafting decisions reliable. Without it, the picker cannot judge why alternatives were rejected.

Can arena candidates write to the same output directory?▼

No. Each candidate must write to its own location, ideally a git worktree or a separate /tmp path. Shared output paths create shared mutable state and cause candidates to overwrite each other.