arena
Have several agents attempt the same task, pick the strongest attempt as the base, and graft the best parts of the others into it. Use for "arena this" or when one attempt at an artifact would lock in the wrong shape.
- License 许可
- MIT
- 所属仓库
- dwin/skills
- 最近更新
- 2026-11-14
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- arena
- description
- Have several agents attempt the same task, pick the strongest attempt as the base, and graft the best parts of the others into it. Use for "arena this" or when one attempt at an artifact would lock in the wrong shape.
- disable-model-invocation
- true
02 资源文件 1
- 参考文档agents/openai.yaml172 B
03 SKILL.md 正文
Arena
Several candidates attempt one task in parallel. You read them all, pick one as the base, and graft into it the few parts of the others that beat it. The output is one artifact.
eval is the neighbor: it measures arms and reports scores. An arena produces
the artifact.
Put the phases below in your task list before you start any candidate: Frame, Fan out, Judge, Pick, Graft, Verify.
Frame
Every candidate gets the same prompt, so the prompt is the contract.
- State the artifact each candidate produces: a module, a design note, a migration plan, a prompt.
- Write the rubric: three to six criteria specific to this task, each one a reader can grade on its own. "Handles a missing config file without a stack trace" is a criterion; "good error handling" is not. The rubric is yours and the judge's. Candidates see only the task.
- Choose the runners. Run at least two. When your host reaches more than
one model family, use more than one: different families fail differently.
Run more when the task has more than one plausible design direction. Use
one model several times when the work is mostly generation and the shape
is not in question.
swarmsays how your host starts agents and which model fits each role. - Give each candidate its own workspace: a worktree through
managing-workspaces, or a temporary directory for an artifact that is not code. - Blind the run by
eval's rules: an ordinary request, a workspace name a user would pick, and no word that the candidate is one of several or is being judged.
Show the frame to the user before the fan-out when they are present.
Fan out
Start every candidate at once. Each gets the task, the shared material it needs, its own workspace path, and one more request: a short note naming the alternatives it considered and why it rejected them.
A candidate that produces nothing drops out. Continue with the rest and note the dropout.
Judge
After every candidate has finished, start one read-only judge on a model from a family other than yours. It sees the rubric and the candidates under neutral labels (A, B, C) and in shuffled order, scores each criterion for each candidate, and recommends a base with its reasons. When no other family is reachable, skip the judge and say so in the result.
Pick
Read every candidate end to end while the judge works. Score each against the rubric one criterion at a time. Then compare with the judge. When you agree on the base, the pick is confirmed. When you disagree, the rubric was ambiguous or one of you is biased: read both sets of reasons before deciding.
Pick the base a future maintainer can extend most easily without breaking its invariants. Between two that feel tied, take the one with the smaller interface or the cleaner boundary.
Graft
Go through each other candidate once more for what beats the base, usually one or two things each. Fold each one in by hand, so the result reads as one design. Pasting blocks across leaves two mental models in one artifact.
When the candidates converge on one shape, that is strong agreement: ship that shape, with no graft needed. When they diverge widely, the frame was under-specified. Fix the frame and run again; averaging the divergence gives an artifact none of them would have made.
Verify
Check the result the way the task is normally checked: its tests, its build,
a run of the real thing, by verifying-claims. When the check finds a problem,
decide where it came from. The frame missed it: fix the frame and run again.
A candidate had it right and you missed the graft: go back to Graft.
When the result is verified, tell the user which candidate workspaces remain
and what each holds. A losing attempt has edits or commits that no kept
branch contains, so removing its workspace deletes that work. Remove them
through managing-workspaces once the user says to discard the attempts.
Result
Deliver the artifact with a short note: the base and why, the judge's verdict or why there was none, each graft and the candidate it came from, what you rejected and why, any convergence or dropout, and the verification result.
04 同仓库更多 Skills
visual-explainer
Render interactive visual explanations, architecture simulators, option tradeoff matrices, permission or blast-radius maps, and phased rollout cards on the host's native visual surface. Use when answering a question that compares three or more options across multiple dimensions, traces a multi-gate request or state flow, maps permissions or blast radius, lays out a multi-phase architecture plan, or when the user asks to visualize or sends `viz`.
writing-for-agents
Writing documents for agents. Use when creating or editing skills, or modifying AGENTS.md or CLAUDE.md.
wizard
Generate an interactive bash wizard that walks a human through steps only they can perform. Use when provisioning infrastructure, setting up credentials or CI secrets, walking an unfamiliar third-party dashboard, or running a one-off migration or cutover. Don't invoke this for steps the agent can perform itself.
why
Find out why code is the way it is from its history, reviews, tickets, docs, and other records, and say how sure each answer is. Use when the user asks why something is built, written, or behaves a certain way, why a choice was made, or where a number or limit came from.
whats-next
Match a task to the skill, flow, or playbook that fits it. Use when the user asks what to do next or which workflow applies, when a task fits no other skill, and for these kinds of work, which have playbooks here - proving a property of concurrent or stateful code with a model checker or verifier, matching a UI pixel for pixel, and diagnosing a live process or a captured profile, trace, or heap snapshot.
wayfinder
Plan a huge chunk of work — more than one agent session can hold — as a shared map of decision tickets on your issue tracker, and resolve them one at a time until the way to the destination is clear.