Conciseness

Nearly every token carries platform-specific knowledge Claude cannot already know (tool names, CLI flags, limits like "20,000 image references per ingest request," exact 402/409/429 error strings, ~7-day expiry). Minor over-explanation keeps it below a 5: the ETL/ELT paragraph ("In pipeline terms: a staged job is the T of an ETL flow...") restates the asset-library-vs-staged distinction already made twice above, and a few caveat sentences could be trimmed. It is not a 3 because no space is spent explaining concepts Claude already knows.

Actionability

Fully executable throughout: copy-paste-ready CLI commands with complete flag lists ("inference rf-cloud data-staging create-batch-of-images --batch-id my-batch --images-dir ./images"), concrete MCP call signatures with named parameters, explicit decision procedures (cpu vs gpu via a timed single-image test run plus Workflow spec inspection), and exact operational limits (batch id charset, format lists, 10,000-entry listing cap). This matches the 5 anchor ("copy-paste ready code or commands; specific examples cover the common cases").

Workflow Clarity

The five-step sequence (install → stage → start → monitor → download) has explicit validation checkpoints for this paid batch operation: "Do not start a job while `pending` is true or `failed` is true," poll-until-terminal guidance, a `retryAfterSeconds` sleep-and-resume loop, `failed` state with logs and a restart hint, and 409 recovery options (inspect, omit optional settings, or pass a new job_id). Feedback loops are present, so the batch-operation cap of 3 does not apply; this matches the 5 anchor ("clear sequence with explicit validation steps; feedback loops for error recovery").

Progressive Disclosure

No bundle files exist, so this is a single-file skill; sections are well-organized and numbered with clear internal navigation ("The full command set for both content types is in sections 1-5") and a clearly signaled, one-level-deep pointer to the separate `cloud-storage` skill for imports. At ~330 lines, however, the full CLI and MCP parameter reference is inlined where the 5 anchor expects content "appropriately split" — e.g., the shared-flag tables in section 3 could live in a reference file. This matches the 4 anchor ("good structure; most content is appropriately placed; minor organization gaps"), not a 3 because nothing is buried or unsignalable.

Specificity

The description lists multiple concrete actions with comprehensive coverage — "running a Roboflow Workflow over many images or videos with batch processing," "Asset Library jobs," "staged jobs (you stage external local/cloud files with the inference-cli)," "monitoring jobs, downloading results, and routing imports to datasources" — which matches the 5 anchor. Per the rubric's voice guideline, the second-person phrasing ("the platform stages workspace images for you," "you stage external local/cloud files") reduces the score by 1. It is not a 3 because it goes well beyond naming 1-2 actions, and not an unpenalized 5 because the second-person constructions are present.

Completeness

It explicitly answers both questions: the when is stated directly ("Use when running a Roboflow Workflow over many images or videos with batch processing") and the what is enumerated concretely (Asset Library jobs, staged jobs via inference-cli, monitoring, downloading results, import routing to datasources). This matches the 5 anchor ("clearly and explicitly answers both what AND when with concrete trigger phrases"), not the 4 anchor where the when could be more explicit.

Trigger Term Quality

Natural terms a Roboflow user would actually say are comprehensively covered: "batch processing," "many images or videos," "Roboflow Workflow," "Asset Library," "monitoring jobs," "downloading results," "datasources (bucket mirror)." No common synonyms or phrasings for this domain are missing, matching the comprehensive-coverage 5 anchor rather than the 4 anchor ("a few natural terms missing").

Distinctiveness Conflict Risk

It occupies a clear niche (Roboflow batch processing) with distinct triggers, and it actively disambiguates the neighboring skill by routing imports away ("routing imports to datasources (bucket mirror) instead"), which minimizes false triggering for import/curation tasks. This matches the 5 anchor ("clear niche with distinct triggers; minimal conflict risk").