Conciseness

The body is dense and operational throughout — no explanations of concepts Claude already knows, no library tutorials — but a few passages could be trimmed: the extended 403-recovery dance in setup step 2, CPU-profile analysis caveats, and rhetorical framing like 'The goal is the truth about the change, not making its author feel good'. This fits the 4 anchor (efficient, minor instances that could be trimmed) better than 5, since not literally every token earns its place; it is well above 3 because there is no genuinely unnecessary explanation.

Actionability

Copy-paste-ready commands throughout — 'node scripts/config.mjs show', 'node scripts/sandbox-gate.mjs --arms cand=<ref>', the full background-launch invocation with the bash -c wrapper and PIPESTATUS handling, 'node scripts/bench-analyze.mjs <runDir>', 'node scripts/bench-db.mjs verify [runId]', 'node scripts/sandbox-sweep.mjs' with flag semantics — plus a verbatim report template. Every referenced script (config.mjs, sandbox-gate.mjs, sandbox-e2e.mjs, sandbox-ssr.mjs, bench-analyze.mjs, bench-db.mjs, bench-status.mjs, bench-collect.mjs, sandbox-sweep.mjs) exists in scripts/. Fully executable and covering the common cases — the 5 anchor.

Workflow Clarity

A clearly sequenced workflow (setup → resolve arms → gate correctness → launch in background → analyze skeptically → report) with explicit validation checkpoints everywhere the operation is costly or destructive: CI-green gating before any VM spend, --dry-run sanity check, 'PASS requires seeing the actual test counts', 'Bench the exact sha the gate prints', fingerprint-header validity check, 'If a gate fails, report the failures and stop', bench-db verify integrity checks, and mandatory confirmation runs for decision-driving claims. Feedback loops for error recovery are a dedicated section. This is the 5 anchor; the batch/destructive cap does not apply because validation is thorough.

Progressive Disclosure

The single reference is real, clearly signaled, and one level deep ('Apply the policy in [references/methodology.md](references/methodology.md)'), and the scripts bundle is consistently referenced by name with per-script roles. However, section 4 both points to methodology.md and then inlines roughly eight bullets of claiming policy anyway, and content like the 403-recovery procedure and profile-diffing caveats arguably belongs in reference files. This fits the 4 anchor (most content appropriately placed, references mostly clear, minor organization gaps) rather than 5, and is well above 3 because what referencing exists is clearly signaled and shallow.

Specificity

Names the domain and multiple concrete actions with named harnesses and metrics: 'paired A/B statistics', 'measured end-to-end through the bench/render-pipeline app (rps, latency, p95; TTFB, RSS and document/Flight bytes)', 'through the react repo's flight-ssr-bench fixture', 'applies correctness gates before measuring', 'reports boot-level confidence intervals'. This matches the 5 anchor (multiple specific concrete actions, comprehensive coverage); it exceeds 4 because there are no coverage gaps in what the skill does.

Completeness

Explicitly answers both what (paired A/B benchmarking of React/Next.js revisions through two named harnesses, with metrics, gates, and CIs) and when ('Use whenever the user asks to bench, perf test, or A/B a React PR... even if they don't say "benchmark"') with concrete trigger phrases. This is the 5 anchor verbatim in structure, not 4, because the when-clause is fully explicit rather than merely present.

Trigger Term Quality

Comprehensive natural trigger terms with synonyms and quoted user phrasings: 'bench, perf test, or A/B', 'is this PR faster', 'does this regress RSC?', 'measure the perf impact of <commit>', plus the catch-all 'any request to quantify a server-side performance difference between two revisions'. Matches the 5 anchor; nothing material a user would naturally say is missing.

Distinctiveness Conflict Risk

A clear niche (paired server-side perf A/B for React/Next.js revisions on Vercel Sandbox VMs) with triggers that would not plausibly fire a general app-run, code-review, or dataviz skill, and an explicit scope boundary ('any request to quantify a server-side performance difference between two revisions belongs here'). Minimal conflict risk — the 5 anchor.