Conciseness

Mostly lean and operational — tool calls with expected outputs, cost/latency notes, and a compact reporting template — with only minor trimmable fat: 'Verify ref_sequence[174] == "R" (Python is 0-indexed, position is 1-indexed)' and 'Python is 0-indexed' restate what Claude already knows, and the opening line 'SAE features are interpretable latent dimensions of the model's hidden state' is light background. Not enough padding to drop to anchor 3's 'noticeably could be tightened'; well above anchor 2.

Actionability

Fully executable end-to-end: concrete tool invocations with parameters and documented return shapes (ESM_explain_variant_mechanism, ESM_score_variant_sae_disruption, ESM_score_variant_sae_batch, ESM_get_sae_features), complete runnable delta-aggregation code, a saturation-variant recipe, and a copy-ready reporting template. Edge behavior is documented ('both tools return a clear error'), covering the common cases like the anchor-5 example.

Workflow Clarity

A clearly sequenced 5-step workflow with an explicit validation checkpoint — 'Verify ref_sequence[174] == "R"' and 'If the reference residue does NOT match, return an explicit error — do not silently mutate the wrong position' — plus error-recovery guidance ('you supplied the wrong isoform / mis-labeled the variant') and a cross-validation checklist for high-stakes calls. The batch/saturation path inherits the same validation, so the batch-validation cap does not apply.

Progressive Disclosure

No bundle files exist (references/, scripts/, assets/ are absent), so everything is inline in a single well-sectioned SKILL.md (~245 lines) with clear headers — 'When to use this skill', 'Required inputs', 'Prerequisites', 'Workflow (5 steps)', 'Interpretation table', 'Honest limitations', 'Cross-validation pattern', 'Reporting format'. Navigation is easy and there is zero reference nesting, but the long-path steps, interpretation table, and cross-validation layers are secondary material that could live in one-level-deep reference files, which keeps it at anchor 4 rather than 5 (which expects content appropriately split across files).

Specificity

Concrete actions are explicit — 'Interpret a missense variant via ESMC-6B Sparse Autoencoder (SAE) feature activations' and 'computes which interpretable SAE features (catalytic, ligand-binding, PTM, structural motif, domain, etc.) are lost or gained at the mutation site' — with the feature categories adding specificity. It stops just short of the 5 anchor's 'multiple specific concrete actions; comprehensive coverage' (essentially two actions), but well above anchor 3's '1-2 concrete actions, not comprehensive' because the enumerated categories and the lost/gained framing make coverage broad.

Completeness

Explicitly answers both: what — 'computes which interpretable SAE features ... are lost or gained at the mutation site'; when — 'Use when standard pathogenicity scores (AlphaMissense, ClinVar) say a variant is damaging but you need a MECHANISTIC explanation'. Both are concrete and use explicit trigger phrasing, matching the anchor-5 example structure.

Trigger Term Quality

Comprehensive natural terms with synonyms: 'missense variant', 'damaging', 'pathogenicity scores', 'AlphaMissense', 'ClinVar', 'LoF', 'mechanistic explanation', and the user-voiced phrase "why is this variant LoF?". These are exactly the phrases a user would say when needing this skill; no significant synonym gaps (score 4's 'a few natural terms missing' doesn't apply).

Distinctiveness Conflict Risk

Clear niche (SAE-based mechanistic explanation of missense variants) plus explicit de-confliction: 'Complements (does not replace) variant-interpretation and variant-to-mechanism skills, which focus on ACMG classification or regulatory mechanism'. This actively routes adjacent requests to sibling skills, minimizing wrong-skill triggering.