Content
82%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A tight, competent body: efficient tables and code, a clearly sequenced remote-compute workflow with an explicit notification checkpoint, and no bundle-file sprawl. The main gaps are a missing minimal `borzoi_run.py` / input-construction example and absent output validation, which hold actionability and workflow clarity just below top marks.
Suggestions
Include a minimal `borzoi_run.py` (or the one-hot input construction) so the remote-job example is fully copy-paste executable.
Add an explicit output validation step, e.g. check `tracks.npz` shape and track count against the expected (B, T, L) format, after `compute_done`.
Show a two-line ref/alt variant-scoring code sketch to make the primary use case (1) in the description concrete.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean throughout — dense prerequisite and troubleshooting tables, minimal code, and gotcha notes like "the base `Borzoi` model has no `targets` attribute" that add non-obvious value while assuming Claude's competence; no section explains concepts Claude already knows. | 5 / 5 |
Actionability | Mostly executable guidance: concrete `from_pretrained` loading, a full `submit_job` call with real arguments, and `attach_job(job_id).result()` for re-entry. However `borzoi_run.py` is passed as a job input without its contents ever being shown, input one-hot construction is only specified as a shape comment, and variant scoring is described ("run ref/alt windows centred on the variant") rather than coded — minor gaps, not fully copy-paste ready. | 4 / 5 |
Workflow Clarity | The remote-compute sequence is clear (read compute_details → submit → "call the `wait_for_notification` brain-tool" → "when the `compute_done` notification arrives, act on its payload" → save_artifacts → attach_job), with the non-blocking checkpoint made explicit. No validation of the output artifact (e.g., checking `tracks.npz` shape or track count) is included, which keeps it below the explicit-validation anchor. | 4 / 5 |
Progressive Disclosure | No bundle files exist and the ~77-line body is cleanly sectioned (Prerequisites, How to run, Output format, Remote compute, Troubleshooting), with orchestration detail appropriately deferred to well-signaled skills (`remote-compute-ssh` / `remote-compute-modal`). It exceeds the under-50-lines simple-skill case and carries the full remote-compute protocol inline, so minor organization gaps remain rather than an ideal split. | 4 / 5 |
Total | 17 / 20 Passed |