Conciseness

The body teaches nothing Claude already knows — every sentence is a repo-specific rule or trap (e.g. ".closingIssuesReferences[0].number drops the repository", "`null` means unobserved, not 'all of it'"). However, stylized prose ("It is gitignored scratch, like every other skill's report; the card is the half that the repository keeps") and some repetition (the no-transcripts case is argued in full in "Before Step 1" and again in Step 1) could be trimmed. Not 5: minor over-elaboration remains; not 3: there is no padding or basic-concept explanation.

Actionability

Copy-paste-ready commands throughout: the three-line environment probe, `bun run --cwd tools/pr-metrics card -- --resolve-issue` (and the `--format markdown` variant), the slug-transformed `git add ".claude/metrics/$(echo "$BRANCH" | tr -c 'a-zA-Z0-9._-' '-').json"`, the append-not-rebuild body edit, the `grep -c '^<details><summary>Session metrics'` guard, and a full RETRO.md template. The judgment steps are backed by concrete signal-to-fix mapping tables. Not 4: the common cases are covered by executable commands with specific flags, including the fallbacks.

Workflow Clarity

Steps 1–7 are clearly sequenced with explicit checkpoints: the pre-run probe ("Three answers, once"), the check-before-append guard with its own failure note ("match the summary line, not the tag: a bare `grep -c '<details>'` then reports a block that is not there"), and per-failure handling ("When a step here cannot run, record which and carry on"). The card is deliberately written first so an early-ending run still leaves the record — an explicit feedback/robustness design. Not 4: validation checkpoints and error-recovery loops are explicit, not implicit.

Progressive Disclosure

Field-level detail is correctly offloaded to `references/card-signals.md` (a real bundle file), clearly signaled in Step 2 ("`references/card-signals.md` maps each of those to the question it answers… Read it before Step 3") — one level deep, no nesting. Repo-level references (`wiki/conventions/session-metrics.md`, `tools/pr-metrics/README.md`) are kept cleanly separate from the run procedure, and the skill names its boundary with `analyse-sessions`. Not 4: the split is appropriate and navigation is easy throughout.

Specificity

Quotes: "write and commit the session-metrics card for the branch", "read what the session actually cost against the difficulty declared before it started", "turn that into concrete changes to the skills, `AGENTS.md`, the wiki, the tests or the brief". Multiple specific concrete actions with named artifacts, covering the skill's full scope. Not 4: there are no minor gaps in the action list — writing, comparing, and converting to changes are all explicit.

Completeness

The description opens with "Use when running the retrospective on a piece of work that has just finished — the retro, normally the step straight after `prep-pr` opens the pull request", an explicit and concrete when-clause, and then states exactly what the skill does (write and commit the card, compare cost to declared difficulty, turn it into concrete changes). Both what and when are explicit with concrete trigger phrases. Not 4: the when is not merely present but specific (timing relative to `prep-pr` and the open PR).

Trigger Term Quality

Quotes: "running the retrospective", "the retro", "prep-pr opens the pull request", "session-metrics card". These are natural phrases a user would say when they need this skill, but common synonyms (e.g. "postmortem", "review the session", "what did this cost") are absent. Not 5: synonym coverage is incomplete; not 3: several genuinely natural trigger terms are present and well-chosen.

Distinctiveness Conflict Risk

Clear niche: the post-PR retrospective slot ("normally the step straight after `prep-pr` opens the pull request") plus an explicit ownership claim — "Also the owner of the card — nothing else writes one carrying the pull request's identity". Minimal conflict risk with other skills; the triggers are specific to this workflow. Not 4: the ownership boundary removes even minor overlap risk.