Conciseness

Every token carries aq- or platform-specific operational knowledge Claude cannot already know ("Many slim images do not have them, including the `pytorch/pytorch` and `nvidia/cuda` runtime tags"; "Every `aq` call prints a one-line `aq: api https://...` banner on stderr; it is not an error"), with zero conceptual padding or pre-taught explanations. Not 4: there is no over-explanation anywhere that could be trimmed.

Actionability

Fully executable, copy-paste-ready guidance at every step: the install one-liner, the base64 script-inlining pattern, a complete `aq job run` invocation, a bounded poll loop with `case "$s" in succeeded|failed|UNSERVABLE|cancelled) break;; esac`, and `aq job pull <job> ./results`. Not 4: the common cases (fine-tuning, HF_TOKEN secret, private registry image) are each covered with concrete, runnable specifics.

Workflow Clarity

Six clearly numbered steps with explicit validation checkpoints and error-recovery feedback loops for every terminal state ("failed: ... Read `aq job logs <job>`, fix the command, and start a new job"; "UNSERVABLE: ... Start it again with more `--gpu-model` choices"), plus a cost-approval gate ("tell the user the GPU models, the hourly price and your estimate of the total, and wait for a yes") and a queued->10min diagnostic path. The batch-operation cap does not apply because status verification, cost reporting, and result reporting are all explicit. Not 4: feedback loops are present for every failure mode.

Progressive Disclosure

Well-organized sections with easy navigation, and bulk material correctly offloaded to clearly signaled external resources ("Full command reference: `aq help`"; "GPU jobs on the CLI: https://docs.aquanode.io/docs/jobs/cli"). However, all ~190 lines of detail — including the `--gpus`/`--disk-gb`/`--any-gpu`/`--gpu-order`/`--secret` flags table and the terminal-state reference — are inlined in SKILL.md with no bundle split. Not 5: content is not split across files and the under-50-line exception does not apply at this length; not 3: the structure is clean and nothing is buried or nested.

Specificity

Concrete actions with comprehensive coverage: "Run a batch job on a cloud GPU with Aquanode's `aq` CLI and bring the results back" plus "from any Docker image and a command" — running the job and retrieving results is the complete capability surface, each stated specifically. Not 4: there are no gaps in coverage of what the skill does.

Completeness

Explicitly answers both parts: the what is "Run a batch job on a cloud GPU with Aquanode's `aq` CLI and bring the results back" and the when is a full "Use when the user wants to run something on a GPU they do not have locally ... for example fine-tuning or training a model, batch inference, transcription, embeddings, evals, benchmarks or rendering" clause. Not 4: the when-clause is fully explicit with concrete trigger examples, matching the anchor-5 exemplar.

Trigger Term Quality

Comprehensive natural vocabulary including synonyms: "fine-tuning or training a model, batch inference, transcription, embeddings, evals, benchmarks or rendering" alongside GPU model names users actually say ("H100, A100, L40S, RTX 4090") and "a GPU they do not have locally". Not 4: no common user phrasing for cloud-GPU work is missing.

Distinctiveness Conflict Risk

"Aquanode's `aq` CLI" and "run something on a GPU they do not have locally" carve a clear niche with distinct triggers that no generic or neighboring skill would claim. Not 4: there is no meaningful overlap risk with closely related skills.