arena

Files (1)

• SKILL.md

SKILL.md (first lines)

# Arena

Fan out N parallel attempts at the same task. Read every candidate end to end. Pick the strongest as the base. Graft the best ideas from the others into it. Verify the synthesized result.

## Start

Open a todolist (the task tools, or a checklist in your reply when the session has none) with one entry per phase before launching anything.

1. Frame
2. Fan out
3. Cross-judge
4. Pick
5. Graft
6. Verify

## Phase A: Frame

The N candidates will receive the same prompt, so the prompt is the contract.

1. State the artifact each candidate is producing.
2. Derive the rubric. State what success looks like for *this* task, then turn it into 3-6 concrete gradeable criteria. The rubric is the picker's tool in Phase D. Candidates only see the task.
3. Pick the runners. Use the `arena runners` line in `~/.claude/rules/pstack-models.md`. If the rule or that line is missing, default to one each on `fable`, `opus`. An `inherit` entry in this line or the cross-judge line means the parent model, so omit `model` for it. If the Agent tool rejects a configured entry, run that seat on `opus` and say so. If it rejects a default, use `opus`. Spawn more when the arena covers multiple design directions. Same model N times when the work is generation-bound rather than judgment-sensitive.
4. Assign output paths. Each candidate writes to its own location (a git worktree via `isolation: "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 `subagent_type: "pstack:worker"` (append the agent suffix from the `# budget:` line of `~/.claude/rules/pstack-models.md`), `run_in_background: true` (false inside a subagent, see below), and `isolation: "worktree"` for worktree output paths, each with the task, the path to the shared grounding, its own output path, and instructions to produce both the artifact and a short rationale. Inside a subagent, spawn with `run_in_background: false`. A subagent is never woken by its own background children, whose results go to the main chat. Several foreground spawns in one message still run in parallel.

A worktree starts at `origin/<default-branch>`, not your checkout. Commit what the candidates need, then tell each in its brief to run `git checkout --detach <sha>` first. Local commits are reachable by SHA from a worktree. Uncommitted changes are not.

Each rationale names the alternatives the candidate considered and what it rejected.

If a candidate fails to produce output, respawn it once. Proceed with N-1 only while two or more remain, and note the dropout in the synthesis record.

If the Agent tool is unavailable (a subagent at the nesting limit), run the steps inline.

## Phase C: Cross-judge

After all Phase B candidates complete, choose one model from the `arena cross-judge pool` line in `~/.claude/rules/pstack-models.md`. If the rule or that line is missing, choose from `fable`, `opus`. Prefer a different model from the parent's. Spawn one `pstack:reader` judge subagent (append the agent suffix from the `# budget:` line of `~/.claude/rules/pstack-models.md`) on that model. It sees the rubric and the candidates by path label, scores each criterion, and recommends a base with rationale. It runs in parallel with the parent's reading in Phase D, not with the candidates themselves. Don't spawn the judge while candidates are still writing.

## Phase D: Pick a base

Read every candidate end to end before picking.

Score each candidate against the rubric criterion by criterion, not on holistic feel. Compare against the cross-judge. 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 on which candidate a future maintainer can extend most easily without breaking invariants. Prefer the cleaner boundary or smaller API when two feel tied, per the Laziness Protocol.

Record the pick and the reason in a short synthesis note alongside the base artifact, including the cross-judge's verdict.

## Phase E: Graft

Walk each losing candidate once more and identify what is worth porting into the base. The signal is 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. Don't paste mechanically. The result has to remain coherent under one mental model.

Record what was grafted, from which candidate, and what was rejected and why.

When N candidates converge on the same shape, that is a strong agreement signal. Note the convergence in the record and ship the consensus shape. No graft is needed. When N candidates wildly diverge, Phase A was under-specified. Reframe and re-run rather than averaging the divergence.

## Phase F: Verify

The synthesized artifact has to hold up under the same scrutiny as any other output, per the **prove-it-works** principle skill.

If verification surfaces a problem the arena did not catch, either Phase A was wrong (re-frame and re-run) or one candidate caught it and you missed the graft (go back to Phase E). Don't paper over.

## Outputs

One synthesized artifact. One short synthesis note alongside, naming the base, the grafts (with source candidate), the rejections, the dropouts if any, and the verification result.

From adnjkm/pstack, MIT.

More from adnjkm/pstack