| name | arena |
| description | Spawn N parallel candidates at the same task, pick a base, graft the strongest parts of the losers into it. Use for /arena, 'arena this', 'throw it in the arena', or when one attempt at a non-trivial artifact would lock in the wrong shape. |
Arena
Fan out N parallel attempts at the same task. Read every candidate end to end.
Pick the strongest as base, graft the best ideas from others, and verify the
synthesized result.
Workflow
1. Frame
Define the task prompt and rubric before spawning:
- State the exact artifact to produce.
- Derive 3-6 concrete gradeable rubric criteria.
- Select distinct model families or subagents for the runner pool.
- Assign isolated output directories or worktrees per candidate.
2. Fan Out
Spawn all N candidate subagents concurrently with the task prompt, shared
context, and instructions to return both the artifact and a brief design
rationale.
3. Cross-Judge
Spawn an independent judge subagent on a distinct model family to score all
candidates against the rubric and recommend a base candidate with rationale.
4. Pick Base
Score each candidate against the rubric criteria. Compare picks with the
cross-judge:
- Pick the candidate that maintainers can extend most easily while preserving
invariants.
- Favor cleaner boundaries and minimal surface area.
- Record the base selection rationale.
5. Graft
Review losing candidates for isolated high-value ideas or edge-case handling:
- Fold improvements into the base cleanly rather than copy-pasting.
- Preserve a single coherent design model.
- Document grafted ideas and rejected alternatives.
6. Verify
Verify the synthesized result against real tests, compilers, or live runtimes.
Output
Deliver the synthesized artifact accompanied by a brief synthesis note detailing
the base selection, grafted components, rejected approaches, and verification
results.
