Conciseness

Phase A step 3 and Phase C are run-on contingency chains ("If the Task tool rejects a configured entry... 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...") that could be substantially tightened or split into a reference. The body never explains concepts Claude already knows, so it is above the 2 anchor, but multiple paragraphs need restructuring rather than minor trims, so it is below 4.

Actionability

Guidance is mostly executable: spawn all N in one message with `run_in_background: true`, output paths like `/tmp/arena-<slug>/candidate-<n>/`, concrete model slugs, and a todolist with named phases. It falls short of 5 because no example candidate prompt, rubric, or synthesis-note template is given, leaving the actual artifacts to be improvised.

Workflow Clarity

Six labeled phases (Frame through Verify) form a clear sequence with a todolist checklist, a cross-judge comparison checkpoint in Phase D, and explicit verification with error-recovery feedback loops in Phase F ("go back to Phase E", "re-frame and re-run"). The batch-operation cap does not apply because validation is explicitly present, and Phase C even specifies timing ("Don't spawn it while candidates are still writing").

Progressive Disclosure

The ~42-line body is organized into clean per-phase sections, but the bulky model-selection fallback policy in Phases A and C is inlined where a one-level-deep reference file would serve better. External references (../../docs/harness/claude-code.md, ~/.cursor/rules/pstack-models.mdc, four bold-named principle skills) are signaled by name only and none exist in this bundle, which is a minor organization gap rather than the clear split of the 5 anchor.

Specificity

"Spawn N parallel candidates at the same task, pick a base, graft the strongest parts of the losers into it" lists multiple specific concrete actions (spawn, pick, graft) that comprehensively cover the skill's capability. It exceeds the 4 anchor because there are no minor gaps in coverage of what the skill does.

Completeness

Both questions are answered explicitly: the what ("Spawn N parallel candidates... pick a base, graft the strongest parts") and the when ("Use for /arena, 'arena this', 'throw it in the arena', or when one attempt... would lock in the wrong shape") with concrete trigger phrases. This matches the 5 anchor; the 4 anchor applies only when the when-clause is less explicit.

Trigger Term Quality

"/arena, 'arena this', 'throw it in the arena'" are precisely the natural phrases a user would say to invoke this, supplemented by the conceptual trigger "when one attempt at a non-trivial artifact would lock in the wrong shape". Coverage is comprehensive rather than leaving a few natural terms missing.

Distinctiveness Conflict Risk

Parallel-candidate generation with base-picking and grafting is a clear niche with distinct 'arena' triggers and minimal conflict risk with typical skills. It is not the 4 anchor because no closely related skill would plausibly capture these triggers.