Conciseness

The body is dense with non-obvious domain knowledge ("median 3.6M tokens against a mean of 15.4M", the `at-open`/`confirmed` exclusion semantics, per-trap decision rules) and never explains concepts Claude already knows. But narrative framing — "This data has already produced two confident findings that were both wrong, and both were wrong the same way", "which is precisely the `cire/host` failure repeated by a different route" — is stylistic emphasis beyond bare instruction: anchor 4 (minor instances that could be trimmed). Not 3, since every section is load-bearing and repo-specific; not 5, since a few flourishes don't strictly earn their tokens.

Actionability

Fully executable, copy-paste-ready commands throughout: `bun run --cwd tools/pr-metrics report -- --coverage/--waste/--json`, an exact `gh pr diff 882 ... | grep -E '^[+-]\s*"[^"]+":\s*"[\^~]?[0-9]'` classifier, a decision table (Small/**High**/High spend = "a hard problem that ended in a one-line fix — leave it alone"), and field-level read rules ("read the report's by line rate (`acted_on_paired - acted_on_paired_by_test`) against `control_changed`"). Anchor 5: commands plus specific decision rules covering the common cases; not 4 because even fallbacks and their caveats (queries.sql, DuckDB local-only) are specified.

Workflow Clarity

A clearly ordered sequence (Step 0 coverage gate → Step 1 run what you need → the six traps → Step 2 report) whose ordering is itself justified, with explicit validation checkpoints: "write them down before continuing", "If nothing is rated, say so at the top of your answer", "report how many you excluded", and a per-claim falsification loop ("the number, the sample size it rests on, and what would have to be true for it to be wrong"). Anchor 5 — explicit gates and feedback loops, not 4's 'most checkpoints present'. The workflow is read-only analysis, so the destructive/batch-missing-validation cap at 3 does not apply.

Progressive Disclosure

A single-file skill (no references/, scripts/, or assets/ bundle exists) with well-organized sections and clearly signaled, one-level-deep pointers to repo docs: "`wiki/conventions/session-metrics.md` holds the schema", "`tools/pr-metrics/README.md` has the commands". Anchor 4 (good structure, minor organization gaps): the six-traps section runs ~50 lines inline, and the closing "What consumes this" provenance section (retro/SessionEnd hook mechanics) is tangential to the run workflow. Not 5 because the >50-line body keeps everything inline and doesn't meet the single-file-5 exception; not 3 because references are clearly signaled and structure is strong.

Specificity

Names several concrete actions tied to real artifacts — "running the report over `.claude/metrics/`", "reporting what is actually supported", "the step before proposing any change to how work is dispatched" — matching anchor 4 (several specific actions, minor gaps). Not 5 because "interpreting it without the traps that have already produced two wrong findings" is abstract: the description never says what interpretation involves (coverage, medians, classification), and not 3 because it clearly enumerates multiple concrete actions rather than 1–2.

Completeness

Explicitly answers both: what ("running the report over `.claude/metrics/`, interpreting it ..., and reporting what is actually supported") and when, with two concrete trigger clauses ("Use when reading the session-performance cards to find where agent effort is going" and "the step before proposing any change to how work is dispatched") — anchor 5. Not 4 because the 'when' is fully explicit rather than 'could be more specific'; the 'Use when...' cap at 3 does not apply.

Trigger Term Quality

"session-performance cards", "where agent effort is going", and "how work is dispatched" are phrases this repo's users would naturally say when they need this skill — good coverage matching anchor 4. Not 5 because common variations are missing ("cost", "spend", "token", "trend" never appear), and not 3 because the terms present are the natural phrasings, not generic jargon.

Distinctiveness Conflict Risk

The niche is clear (corpus-level session-metrics analysis gating dispatch changes), but the sibling `retro` skill — named only in the body — covers the adjacent trigger of reading a session card, and the description does not say corpus-vs-one-card. That is anchor 4 (mostly distinct, minor overlap risk with a closely related skill). Not 5 because that closely related skill exists and the line is not drawn here; not 3 because the niche is far more specific than generic domain language.