Content
68%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 lean, highly actionable reference whose code recipes are near copy-paste ready and whose gotchas (padding, BOS/EOS, TORCH_HOME, package disambiguation) are exactly the non-obvious knowledge worth a skill. The main deductions are the absence of validation steps in the batch-embedding and remote-job workflows and a confusing, partially duplicated remote-compute prose tail.
Suggestions
Add explicit verification to the batch workflow — e.g., assert `seq_embs.shape == (B, D)` and that `len(residue_embs)` matches the input sequence count before using the embeddings, turning the batch recipe into a validate-then-proceed loop.
Complete the masked-LM example: show the actual masking step (`toks[0, pos] = alphabet.mask_idx` or equivalent) and the `logit[mut] − logit[wt]` comparison as runnable code instead of a comment hint.
Trim the remote-compute prose (the `suppressed`/`committed` sentences) to a one-line delegation to the `remote-compute-ssh` skill, keeping only the submitJob snippet and the TORCH_HOME note.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and earned — executable snippets, a prerequisites table, a model table, and genuinely non-obvious gotchas (BOS/EOS/padding slicing) with no filler explaining what proteins or transformers are. However, the remote-compute tail ("A final `.result()` read reports whether its follow-up was `suppressed` or had already been `committed`...") is jargon-heavy, confusing, and could be trimmed or delegated to the remote-compute-ssh skill it already points to. | 4 / 5 |
Actionability | Embedding (single and batch), contact prediction, and job submission are copy-paste ready with exact slice indices and shape annotations. Two gaps keep it from 5: the masked-LM snippet is a fragment reusing prior variables and only hints ("mask the position and compare logit[mut] − logit[wt]") without showing how, and `environment=...` is an unexplained placeholder. | 4 / 5 |
Workflow Clarity | Sections are logically ordered (prerequisites → recipes → output format → remote compute → troubleshooting) and the compute flow has checkpoints (save job_id, non-blocking status/result queries), but batch operations lack explicit validation — the batch-embedding snippet has no verification (e.g., shape/ordering checks) and the remote 200-sequence job has no result verification. Per the rubric, batch workflows missing validation steps cap workflow clarity at 3. | 3 / 5 |
Progressive Disclosure | No bundle files exist; the single ~130-line body is well-sectioned and correctly delegates remote-compute and structure-prediction detail to the `remote-compute-ssh` and `esmfold2` skills (clear one-level cross-skill pointers). It is not the under-50-line case, and the remote-compute section partially duplicates material that belongs to the referenced skill, so it fits "good structure; minor organization gaps" rather than 5. | 4 / 5 |
Total | 15 / 20 Passed |