Arena
N parallel candidates attack the same brief; the best artifact is picked
criterion by criterion and strengthened by grafting. Diversity of candidates
is the mechanism: agreement across independent attempts is high-signal, and
divergence exposes what the brief under-specified.
When to use
- a design or implementation has more than one defensible shape, and the
first attempt would lock in the wrong one
- judgment-sensitive work: a migration strategy, an API surface, a data
structure choice, a subtle algorithm
- the user asks for a bakeoff, a comparison of approaches, or "try a few
ways and pick"
When not to use
- mechanical edits, contained one-file changes, or work with an obvious
single answer — run
execute instead
- research questions whose deliverable is an answer, not an artifact — run
research
Phase A — Frame
Every candidate gets the same brief, so the brief is the contract.
- State the artifact each candidate produces.
- Derive the rubric: 3-6 concrete, gradeable criteria for this task. The
rubric is the picker's tool in Phase D; candidates never see it.
- Pick the runtime:
- In-session: parallel
Task dispatches to oh-my-claudecode:executor,
one per candidate, each given its own output location.
- Multi-provider: seat candidates for model diversity — read
team.roleRouting from the project config (.claude/omc.jsonc) or the
user config (claude-omc/config.jsonc; omc config prints the
effective values), collect up to three distinct provider families in
canonical role order, and fill remaining seats from the default
families — claude, codex, gemini — not already seated; with no routing
config, use
OMC_TEAM_WORKTREE_MODE=detached omc team --no-decompose 1:claude,1:codex,1:gemini "<brief>".
Worktree isolation must be on — an unset mode resolves to disabled
and all candidates would share one checkout — and --no-decompose
keeps every candidate on the same brief instead of splitting a
multi-point brief across workers (one writer per worktree).
Use diverse models when the work is judgment-sensitive; the same model N
times when it is generation-bound.
- Give each candidate an exclusive output location — a git worktree when
possible, otherwise a scratch directory. Candidates must not share
writable state.
After starting a multi-provider team, wait for every worker to reach a
terminal state and collect its artifact. Keep the worktrees alive through
judging and grafting; run team shutdown/cleanup only in the final cleanup
step after verification.
Phase B — Fan out
Spawn all candidates in one message, in the background, each with the brief,
its output location, and instructions to return the artifact plus a short
rationale naming the alternatives it considered and rejected. If a candidate
produces nothing, proceed with N-1 and note the dropout. Keep at least two
candidates: when dropouts leave fewer, spawn replacements — each on a
remaining candidate's model, or on the parent's model when none remain —
until two produce artifacts, and say so.
Phase C — Cross-judge
After all candidates finish, spawn one read-only judge on a model from a
provider family none of the candidates ran on. If that dispatch is rejected,
fall back down the remaining cross-family models; when no cross-family
dispatch runs, skip the judge, pick the base alone in Phase D, and say so.
If you cannot tell which family the candidates ran on, try the models you
would have picked and name in the verdict that the family check did not run.
The judge sees the rubric and the candidates by path label, scores each
criterion, and recommends a base with rationale. It runs while you read.
Phase D — Pick
Read every candidate end to end. Score each against the rubric criterion by
criterion, not on holistic feel, and compare with the cross-judge when
Phase C ran one. Agreement
on the base confirms the pick; disagreement means one of you is biased or the
rubric is ambiguous — read both rationales before deciding. Tie-break toward
the smaller API and the cleaner boundary (minimal-code-discipline).
Phase E — Graft
Walk each losing candidate once more for what is worth porting — usually one
or two things, not most of it. Fold grafts in by hand so the result stays
coherent under one mental model. Candidates that converge on one shape are
strong agreement: note it and ship the consensus shape. Candidates that
wildly diverge mean Phase A was under-specified: reframe and rerun rather
than averaging the divergence.
Phase F — Verify
Verify the synthesized artifact under the verify skill protocol — real
command output, fresh evidence. A problem the arena missed means either the
rubric was wrong (reframe and rerun) or a graft was missed (back to Phase E).
Do not paper over it.
Rules
- Never a default gate: arena runs only when explicitly invoked.
- The lead session owns the result; candidate and judge outputs are inputs,
not verdicts.
- A later
review judges the synthesized artifact, not the candidates.
1---2name: arena3description: "Run N parallel candidates at one brief and synthesize: rubric-first framing, isolated fan-out, a cross-judge on a different model, criterion-by-criterion pick, hand grafting. When one attempt would lock in the wrong shape."4aliases: []5metadata:6 source: https://skillmd.com/skills/tuyv/arena7 credits:8 arena-workflow: pstack9 author: Lauren Tan10 url: "https://github.com/cursor/plugins/tree/main/pstack"11 port: pstack-claude12 port-author: Michael Denyer13 port-url: "https://github.com/michael-denyer/pstack-claude"14---1516# Arena1718N parallel candidates attack the same brief; the best artifact is picked19criterion by criterion and strengthened by grafting. Diversity of candidates20is the mechanism: agreement across independent attempts is high-signal, and21divergence exposes what the brief under-specified.2223## When to use2425- a design or implementation has more than one defensible shape, and the26 first attempt would lock in the wrong one27- judgment-sensitive work: a migration strategy, an API surface, a data28 structure choice, a subtle algorithm29- the user asks for a bakeoff, a comparison of approaches, or "try a few30 ways and pick"3132## When not to use3334- mechanical edits, contained one-file changes, or work with an obvious35 single answer — run `execute` instead36- research questions whose deliverable is an answer, not an artifact — run37 `research`3839## Phase A — Frame4041Every candidate gets the same brief, so the brief is the contract.42431. State the artifact each candidate produces.442. Derive the rubric: 3-6 concrete, gradeable criteria for this task. The45 rubric is the picker's tool in Phase D; candidates never see it.463. Pick the runtime:47 - **In-session**: parallel `Task` dispatches to `oh-my-claudecode:executor`,48 one per candidate, each given its own output location.49 - **Multi-provider**: seat candidates for model diversity — read50 `team.roleRouting` from the project config (`.claude/omc.jsonc`) or the51 user config (`claude-omc/config.jsonc`; `omc config` prints the52 effective values), collect up to three distinct provider families in53 canonical role order, and fill remaining seats from the default54 families — claude, codex, gemini — not already seated; with no routing55 config, use56 `OMC_TEAM_WORKTREE_MODE=detached omc team --no-decompose 1:claude,1:codex,1:gemini "<brief>"`.57 Worktree isolation must be on — an unset mode resolves to `disabled`58 and all candidates would share one checkout — and `--no-decompose`59 keeps every candidate on the same brief instead of splitting a60 multi-point brief across workers (one writer per worktree).61 Use diverse models when the work is judgment-sensitive; the same model N62 times when it is generation-bound.634. Give each candidate an exclusive output location — a git worktree when64 possible, otherwise a scratch directory. Candidates must not share65 writable state.66 After starting a multi-provider team, wait for every worker to reach a67 terminal state and collect its artifact. Keep the worktrees alive through68 judging and grafting; run team shutdown/cleanup only in the final cleanup69 step after verification.7071## Phase B — Fan out7273Spawn all candidates in one message, in the background, each with the brief,74its output location, and instructions to return the artifact plus a short75rationale naming the alternatives it considered and rejected. If a candidate76produces nothing, proceed with N-1 and note the dropout. Keep at least two77candidates: when dropouts leave fewer, spawn replacements — each on a78remaining candidate's model, or on the parent's model when none remain —79until two produce artifacts, and say so.8081## Phase C — Cross-judge8283After all candidates finish, spawn one read-only judge on a model from a84provider family none of the candidates ran on. If that dispatch is rejected,85fall back down the remaining cross-family models; when no cross-family86dispatch runs, skip the judge, pick the base alone in Phase D, and say so.87If you cannot tell which family the candidates ran on, try the models you88would have picked and name in the verdict that the family check did not run.89The judge sees the rubric and the candidates by path label, scores each90criterion, and recommends a base with rationale. It runs while you read.9192## Phase D — Pick9394Read every candidate end to end. Score each against the rubric criterion by95criterion, not on holistic feel, and compare with the cross-judge when96Phase C ran one. Agreement97on the base confirms the pick; disagreement means one of you is biased or the98rubric is ambiguous — read both rationales before deciding. Tie-break toward99the smaller API and the cleaner boundary (`minimal-code-discipline`).100101## Phase E — Graft102103Walk each losing candidate once more for what is worth porting — usually one104or two things, not most of it. Fold grafts in by hand so the result stays105coherent under one mental model. Candidates that converge on one shape are106strong agreement: note it and ship the consensus shape. Candidates that107wildly diverge mean Phase A was under-specified: reframe and rerun rather108than averaging the divergence.109110## Phase F — Verify111112Verify the synthesized artifact under the `verify` skill protocol — real113command output, fresh evidence. A problem the arena missed means either the114rubric was wrong (reframe and rerun) or a graft was missed (back to Phase E).115Do not paper over it.116117## Rules118119- Never a default gate: arena runs only when explicitly invoked.120- The lead session owns the result; candidate and judge outputs are inputs,121 not verdicts.122- A later `review` judges the synthesized artifact, not the candidates.