Conciseness

The body runs ~778 lines and is padded with incident narratives and verification transcripts: "Cost a staging gate run on 2026-08-28", "left the runner down for about seven minutes mid-gate on 2026-09-10", "session 58ce3a58-8d04-40ac-99e4-c44eaa5d7b06", "Verified PASS 3/3 runs after the tunnel-seat fix (2026-08-06)", plus trial-by-trial model-behavior transcripts and open editorial questions ("Worth a call: whether the guidance text needs to be more directive"). Not 1 because none of it explains concepts Claude already knows — it is all non-obvious project knowledge; not 3 because the volume of narrative padding goes well beyond "some unnecessary explanation", and dated/time-sensitive facts are scattered throughout rather than confined to a changelog or old-patterns section.

Actionability

Quotes: "uv run resources/qa_product.py --all --custom-slug <vault-slug> --custom-name \"<display-name>\" --require-store" (plus single-cell and single-journey variants), the complete executable /tmp/replace-runner.sh hook script with kill/start/health-wait, "uv run resources/session_control.py --cells all --harness pi_core --sandbox local", and enumerated AGENTA_QA_* env vars with defaults ("--burst-size (default 16)"). Copy-paste-ready commands cover the common cases: full run, one cell, one journey, continuity, session control, and commit approval.

Workflow Clarity

Quotes: "EXPORT the three variables, do not just set them", "Any FAIL blocks the release until triaged", "the release is not green until each of those has a recorded result of its own", "Run dead LAST", and explicit SKIP-vs-FAIL semantics ("A SKIPped integration test in a security- or concurrency-bearing area is a FAILURE in your summary line"). The run sequence is clear with strong validation checkpoints and error-recovery guidance throughout. Not 5 because the full release-run ordering across the multiple drivers (qa_product.py, session_control.py, matrix cells, qa_commit_approval.py) is distributed across sections and deferred to coverage.md rather than given as one consolidated checklist; not 3 because checkpoints are explicit and pervasive, and destructive operations (kill hooks) carry explicit validation.

Progressive Disclosure

A "Resources (read on demand)" section signals one-level-deep references with purpose descriptions ("resources/coverage.md — the cells (harness × sandbox × auth), the journeys... a table of what each cell needs beyond the three env vars"), and "Full runbook and ordering: resources/coverage.md" points elsewhere for detail. However, 400+ lines of per-cell documentation, incident history, and evidence transcripts that belong in coverage.md/LESSONS.md are inlined in SKILL.md itself. Not 4 because a large fraction of the content is inappropriately placed in the overview file rather than having "minor organization gaps"; not 2 because structure is present and references are clearly signaled, not buried. Note also that the referenced resources/ files are not present in this bundle, so the references could not be verified against actual files.

Specificity

Quotes: "Drives the same product endpoint the playground drives", "asserts on the SSE frame stream and real side effects, never on model prose", "works against any deployment (cloud or self-hosted) from three env vars" — several concrete, specific actions are named. It is not 5 because the opening "Run the agent release gate" is self-referential (names the skill rather than a capability) and the action coverage rests on 2-3 mechanism statements rather than a comprehensive enumeration; it is well above 3 because the actions listed are concrete and multiple.

Completeness

Quotes: the what is concrete ("a portable, wire-level QA harness for the agent runtime... asserts on the SSE frame stream and real side effects") and the when is explicit with concrete trigger phrases ("Use before an agent-workflows release, or after changing the runner..." plus the quoted trigger list). This directly matches the anchor "clearly and explicitly answers both what AND when with concrete trigger phrases".

Trigger Term Quality

Quotes: "run the release gate", "QA the agent runtime", "does the agent still work end to end", "pre-release agent QA" — four natural phrases a user would actually say, plus change-event triggers ("after changing the runner, the SDK agent adapters, the runner Docker images, or the agent service"). Not 5 because common synonyms such as "smoke test" or "regression check" are absent; clearly above 3 because four explicit natural trigger phrases plus event-based conditions exceed "some relevant keywords".

Distinctiveness Conflict Risk

Quotes: "agent release gate", "QA the agent runtime", "pre-release agent QA" — a clear niche (release QA for the agent runtime) with distinct, unambiguous triggers. Minimal overlap risk with other skills; a user saying these phrases means this skill.