Conciseness

The body is lean and imperative (~25 lines) with no explanation of concepts Claude already knows; the only trimmable material is the brief incident-rationale intro ("most incidents start with a command aimed at the wrong one") and the Rules section substantially restating workflow steps 3-6. Not 5 because that Workflow/Rules duplication could be tightened; not 3 because there is no padding or unnecessary conceptual explanation.

Actionability

Fully concrete and executable: exact env vars and files to read (`CONVEX_DEPLOYMENT` in .env.local, `convex.json`), exact commands (`npx convex run --prod`, `npx convex env list`), exact MCP flags (`--deployment dev`, `--cautiously-allow-production-pii`, `--dangerously-enable-production-deployments`, `--disable-tools run,envSet,envRemove`), a concrete classification taxonomy, and a copy-paste announcement format (`target: dev (joyful-capybara-123, personal dev)`). Not 4 because the enumerated commands and flags cover the common deployment-affecting cases with no significant gaps.

Workflow Clarity

Clear numbered sequence (IDENTIFY → ANNOUNCE → consent → MCP defaults → read-only → diagnosis → ambiguity) with explicit validation checkpoints ("If two sources disagree, resolve before proceeding"; "Ambiguity = stop... never guess") and a genuine error-recovery feedback loop (a deploy that "didn't change anything" means re-run step 1, not re-deploy). Not 4 because both the checkpoints and the recovery loop are present for this destructive-operation context, satisfying the feedback-loop requirement.

Progressive Disclosure

The skill is under 50 lines, needs no external references (no references/, scripts/, or assets/ files exist in the bundle), and is well organized into a Workflow section and a Rules section with no dangling references. Per the rubric's guideline for short self-contained skills, well-organized sections alone earn the top score; not 4 because there is no content that belongs in a separate file.

Specificity

Lists several concrete actions — "Classify + announce the target Convex deployment before any deployment-affecting command", "fresh explicit consent for prod actions", "session read-only mode" — anchored to a specific domain. Not 3 because it names more than 1-2 actions; not 5 because notable behaviors from the body (MCP prod-flag risk splitting, wrong-deployment diagnosis) are absent, leaving minor gaps in coverage.

Completeness

The 'what' is clear (classify, announce, consent-gate, read-only) and a 'when' is explicitly stated ("before any deployment-affecting command"), so it exceeds the weakly-implied level of 3. Not 5 because there is no 'Use when...' clause or concrete user-trigger phrasing (e.g., 'Use when the user asks to deploy to Convex or run convex commands against prod'), so the when could be more explicit.

Trigger Term Quality

Contains the natural terms a user would say — "Convex", "deployment", "prod", "read-only" — plus the "deployment-affecting command" framing. Not 5 because common synonyms and variations ("production", "npx convex deploy", "preview", "dev") are missing; not 3 because the core natural vocabulary is present, not just technical jargon.

Distinctiveness Conflict Risk

Occupies a clear niche — guarding "the target Convex deployment" with prod consent-gating and read-only session semantics — that no generic skill description would collide with. Not 4 because even closely related Convex deployment skills would not trigger for this description's specific guard framing; conflict risk is minimal.