| name | arena |
| description | Fan out N blind candidates at one task, pick the strongest as the base, graft the best ideas from the losers into it, verify the result. Use when one attempt at a non-trivial artifact — a design, a tricky function, a doc — would lock in the wrong shape. |
| argument-hint | [the artifact to produce] |
Arena
One attempt commits to one shape. When the shape is the risky part, produce
several and choose.
When
A non-trivial artifact whose first draft sets a structure everything later
inherits: a design, a tricky function, a document. Skip it when the shape is
already settled.
Protocol
- Frame — write the task once, exactly as every candidate will receive
it, plus the acceptance test the winner has to pass. The prompt is the
contract; a vague one produces N incomparable artifacts.
- Fan out — spawn N fresh-context subagents (default 3) in one message,
all with the same task, each writing to its own path. They stay blind to
each other. Each returns the artifact and a short rationale naming what it
considered and rejected.
- Read — read every candidate end to end. Skimming picks the most
familiar surface, not the strongest one.
- Pick a base — the candidate a maintainer can extend most easily.
Between two that feel tied, take the smaller surface.
- Graft — port the best ideas from the losers into the base by hand,
usually one or two per candidate. The result stays coherent under one
mental model.
- Verify — run the repo's checks, or the acceptance test from step 1,
against the synthesized artifact.
Candidates that converge on one shape are agreement: ship the consensus and
skip the graft. Candidates that wildly diverge mean step 1 was
under-specified: reframe and re-run rather than averaging them.
Report
Which candidate won and why. What was grafted from which loser, and what was
rejected. The verification evidence.