Mostly efficient and free of basic-concept filler (it does not explain what a PDF/MWU/Spearman test is), but at ~440 lines it includes meta-commentary ('SAE is shown as the worked example because the surrounding skills in this collection are SAE-themed') and discursive prose in 'Honest limitations' that could be tightened. It is not score 3 because not every token earns its place; it is not score 1 because it does not pad with concepts Claude already knows.

Packed with complete, executable Python — sae_drop_per_variant, categorize, the NaN-coverage gate, the Spearman sign check, mannwhitneyu, the robustness sweep, and the matplotlib visualization are all copy-paste ready with real signatures and return shapes. It is not score 2 because the guidance is concrete code rather than pseudocode or vague description (only the option-C log-odds line is slightly under-specified).

Clear numbered sequence (Steps 1→2→3→3.5→4→5→6) with an explicit MANDATORY validation gate (the pre-MWU NaN sanity check that raises on insufficient coverage) and feedback loops ('If you hit the ValueError: debug the predictor computation step…') plus a robustness sweep. It is not score 2 because validation checkpoints are explicit and error-recovery guidance is stated, not merely implied.

Well-organized with clear headers and a cross-references table, but it is a long single-file doc with the five predictor options (notably the AlphaMissense bin-parsing block) inlined rather than split into one-level-deep reference files; no bundle files exist in references/scripts/assets. It is not score 3 because for a skill this size SKILL.md is not an overview pointing to detailed materials, and not score 1 because it is not a monolithic wall of text and has no deeply nested references.

Lists multiple specific concrete actions — 'Computes per-variant predictor scores, splits variants into neutral vs disruptive groups by DMS effect, runs a Mann-Whitney U test on the predictor scores, and sweeps the stratification thresholds for robustness' — matching the score-3 anchor. It is not score 2 because it goes well beyond naming a domain plus a couple of actions; it enumerates a full concrete pipeline. The single generic 'you' in the trigger clause follows the standard 'Use when…' pattern of the rubric's own good examples rather than the 'You can use this' framing the voice penalty targets, so no deduction is applied.

Explicitly answers both what and when: the action list states what it does, and 'Use when you need to know whether a predictor's scores track real functional disruption on a specific protein' gives an explicit trigger. It is not score 2 because the 'when' is stated outright rather than merely implied.

Good coverage of natural terms the target audience would actually say — 'AlphaMissense', 'ESM-C SAE', 'ESM logits', 'EVE', 'conservation scores', 'deep mutational scanning (DMS)', 'variant-effect predictor'. It is not score 2 because it surfaces the common predictor names and the DMS acronym users would invoke, not just one generic keyword.

Clear niche — benchmarking/validating a variant-effect predictor against DMS data — with distinct triggers unlikely to fire for adjacent skills; the body's 'Not for' section further separates it from residue-interpretation and single-variant skills. It is not score 2 because the DMS-validation framing is specific and unlikely to overlap with sibling skills.