/pstack:arena
AutoIn one lineSeveral AIs each try the problem, another company's model scores them, and the best one absorbs the others' good ideas.
- When to use
- Choices that are hard to change later, like names, formats or algorithms, where you want options.
/pstack:arena This cache key format will be hard to change later. Give me 5 options.Just asking vs. using it
The same request, typed plainly or with /pstack:arena. Here's what changes.
With /pstack:arenaWhat the document says
7 sections, about 900 words, 4 min to read. Click a section to read it.
OpeningIt opens a todolist with one entry per phase before launching anything.
- 1Phase A: FrameStates what each candidate produces, sets a 3 to 6 point rubric, picks runners, and gives each its own worktree.
- 2Phase B: Fan outSpawns all runners in one message. Each returns its work and a rationale naming what it rejected.
- 3Phase C: Cross-judgeAfter all finish, one judge from another model family scores each criterion without editing and recommends a base.
- 4Phase D: Pick a baseReads every candidate end to end, scores each criterion, and compares with the judge. Disagreement means rereading both rationales.
- 5Phase E: GraftHand-ports one or two good ideas from each loser into the base. Wild divergence means reframing and rerunning.
- 6Phase F: VerifyChecks the result per prove-it-works. A miss sends it back to Frame or to Graft.
- 7OutputsDelivers one artifact plus a note: the base and why, the judge's verdict, each graft's source, rejections, and verification.
Wording worth borrowing
A few sentences from the original: what each one makes the AI do, why it works, and how you can borrow it.
Write the prompt as a contract
Every candidate gets the same prompt, so the prompt is the contract.
- Makes the AI
- It writes the task and scoring rules before sending anyone, so the versions can be fairly compared.
- Why it works
- Calling it a contract reminds the AI that anything left out can't be held against a candidate later, so it writes it all down.
- Borrow it
- When several AIs or people do the same job, give one shared brief: "One page, under 800 words, three real examples."
Score against a rubric
Score each against the rubric criterion by criterion, not on holistic feel
- Makes the AI
- When picking a base it shows each version's score on each point, so you can see why it chose one.
- Why it works
- Overall feel gets swayed by polish or length; scoring point by point splits one fuzzy call into small checkable ones.
- Borrow it
- When comparing options, give the rubric: "Score speed, cost and ease of change 1-5 each, then pick."
Bring in an outside judge
The judge is never from the parent's family
- Makes the AI
- Candidates are scored by a model from another family, so no model gets to favor its own style.
- Why it works
- Naming where bias comes from and ruling it out with a hard rule works better than asking for objectivity.
- Borrow it
- To review AI work, switch reviewers: "Claude wrote this; have GPT critique it as a reader."
Trace bad results to the question
Reframe and re-run rather than averaging the divergence.
- Makes the AI
- When versions differ wildly it restates the task and reruns, instead of stitching them into a mess.
- Why it works
- It reads messy results as a badly framed question, so the AI fixes the source instead of splitting the difference.
- Borrow it
- When AI answers contradict, don't ask for a compromise; ask: "What was unclear in my request that led to such different answers?"
Full text
The SKILL.md the AI reads, as written.
descriptionSpawn N parallel candidates at the same task, pick a base, graft the strongest parts of the losers into it. Use for /pstack:arena, 'arena this', 'throw it in the arena', or when one attempt at a non-trivial artifact would lock in the wrong shape.
Arena
Open a todolist with one entry per phase (Frame, Fan out, Cross-judge, Pick, Graft, Verify) before launching anything.
Phase A: Frame
Every candidate gets the same prompt, so the prompt is the contract.
- State the artifact each candidate is producing.
- Derive the rubric, 3-6 concrete gradeable criteria for what success looks like on this task. It is the picker's tool in Phase D. Candidates only see the task.
- Pick the runners. Use the
arena runnersline in the pstack settings file (~/.cursor/rules/pstack-models.mdcin Cursor,~/.claude/rules/pstack-models.mdin Claude Code,~/.agents/pstack-models.mdin other harnesses). If the file or that line is missing, default to one each onclaude-opus-5-5-xhighandgrok-4.7-xhigh-fast. Anautoorinherit-parententry in this line means the parent model, so omitmodelfor it. If your subagent tool rejects a configured entry, run that seat on its family's default and say so. Families go by prefix:claude-*, andgrok-*withcursor-grok-*andcomposer-*. Any other prefix, likegpt-*, is its own family. A refusedcomposer-*entry runs ongrok-4.7-xhigh-fast. If the family has no default, or your subagent tool rejects that default, use the closest valid slug of the same family from its error message or your harness's model list, or drop the seat and say so when there is none. If dropped seats leave fewer than two runners, add runners until there are two, or one when the line has one entry. Run each added runner on a remaining runner's model, or omitmodelso it runs on the parent model when none remain, and say so. Spawn more runners when the arena covers multiple design directions. Use the same model N times when the work is generation-bound rather than judgment-sensitive. Pick each subagent's reasoning level and fast or standard speed for the job, per the Subagents section of the poteto-mode skill. - Give each candidate its own output location (a git worktree where possible, otherwise
/tmp/arena-<slug>/candidate-<n>/), per the separate-before-serializing-shared-state principle skill.
Phase B: Fan out
Spawn all N subagents in one message with run_in_background: true, each with the task, the path to the shared grounding, its own output path, and instructions to produce the artifact and a short rationale naming the alternatives it considered and what it rejected.
If a candidate produces no output, proceed with N-1 and note the dropout.
Phase C: Cross-judge
After all Phase B candidates complete, choose one model from the arena cross-judge pool line in the pstack settings file. If the file or that line is missing, choose from claude-opus-5-5-xhigh and grok-4.7-xhigh-fast. Spawn one readonly judge subagent. The judge is never from the parent's family, as Phase A step 3 defines families. Try only the entries from another family, in pool order. An auto or inherit-parent entry is the parent model, so it never judges. With the default pool, a Claude parent gets grok-4.7-xhigh-fast. A parent from any other family gets claude-opus-5-5-xhigh first. If your subagent tool rejects an entry, use a valid slug of the same model at the closest level or speed from the error message or your harness's model list, or move to the next entry when there is none. If no entry from another family runs, try the default pool's entries from another family the same way. If none of those runs, skip the judge, pick the base alone in Phase D, and say so. If you do not know the parent's model, try the pool's real slugs in pool order and say 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 in parallel with your reading in Phase D. Don't spawn it while candidates are still writing.
Phase D: Pick a base
Read every candidate end to end before picking. 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 was ambiguous. Read both rationales before deciding.
Pick the base a future maintainer can extend most easily without breaking invariants. When two feel tied, prefer the cleaner boundary or smaller API, per the Laziness Protocol.
Phase E: Graft
Walk each losing candidate once more for what is worth porting into the base, usually one or two things per candidate, not most of it. Fold each graft in by hand, per the redesign-from-first-principles principle skill, so the result stays coherent under one mental model. Don't paste mechanically.
If the candidates converge on one shape, that is strong agreement. Note it and ship the consensus shape. No graft is needed. If they wildly diverge, Phase A was under-specified. Reframe and re-run rather than averaging the divergence.
Phase F: Verify
Verify the synthesized artifact like any other output, per the prove-it-works principle skill. If verification finds a problem the arena missed, either Phase A was wrong (re-frame and re-run) or a candidate caught it and you missed the graft (go back to Phase E). Don't paper over it.
Outputs
One synthesized artifact, with a short synthesis note alongside naming the base and why, the cross-judge's verdict or that Phase C skipped the judge, each graft with its source candidate, what was rejected and why, any convergence or dropouts, and the verification result.
Original by Lauren Tan (poteto), MIT license. Checked against the pstack mirror at 97de40d. View on GitHub