Conciseness

The body is dense and assumes competence — no padding about what drugs or dose-response are, and the model-selection logic is delivered via compact tables ('You measured... | Use model | Tool | Input'). It falls short of anchor 5 because of minor duplication and mild over-explanation: the Chou-Talalay CI interpretation ('CI<1 synergy, =1 additive, >1 antagonism') appears nearly verbatim in both Step 1 and Step 3, and the Step 1 null-model theory (E_a + E_b - E_a*E_b) is knowledge Claude largely already has. Not 3 because every section still carries operational value and nothing reads as filler.

Actionability

Concrete and executable: exact tool names with full input parameter listings in the Step 0 table ('effect_a, effect_b, effect_combination (each a fraction 0-1)', 'doses_a, doses_b, viability_matrix'), a copy-paste bash command with a worked example and expected output ('tu run DrugSynergy_calculate_bliss ... -> expected 0.58, bliss_synergy_score 0.12'), and a runnable script with usage. Not 5 because only one of the five tools gets a complete runnable command; the other four are specified by parameter name only, so the common Loewe/CI/ZIP invocations are not copy-paste ready.

Workflow Clarity

A clean numbered sequence (Step 0 pick model by data -> Step 1 understand the null -> Step 2 run -> Step 3 interpret -> Step 4 gotchas) with explicit checkpoints: the scale-conversion gate in Step 0 ('Effects must be on a consistent inhibition scale... convert: inhibition = 1 - viability/100'), the error signal that 'the tools say so' when the Hill fit fails on <3 dose points, and a gotchas checklist covering ceiling effects, model shopping, and score-vs-efficacy. The calculations are non-destructive, so no validate/retry loop is required. Not 4 because validation checkpoints are explicit and woven into the flow rather than merely implied.

Progressive Disclosure

The SKILL.md is a compact overview with well-labeled sections and two clearly signaled, one-level-deep bundle pointers: 'scripts/synergy_reference.py' (verified to exist, and its docstring matches the body's description of it) and a 'Related skills' list routing to sibling skills for curve fitting and pre-computed lookups. No nested references, no content that clearly belongs in a separate file — the inline tables are short enough to justify inlining for a skill this size. Not 4 because navigation is unambiguous and every pointer resolves to a real file with a stated purpose.

Specificity

The description lists multiple concrete actions — 'quantify whether two drugs together are synergistic, additive, or antagonistic', 'Explains which model to use, what data each one needs, and how to read the score' — and comprehensively names all five reference models (Bliss, HSA, Loewe, ZIP, Chou-Talalay CI). Coverage of the skill's capabilities is complete with no generic filler. Not 4 because nothing meaningful is missing: the what, the models, the data inputs, and the output (a synergy score) are all explicit.

Completeness

Both halves are explicit: the what is 'quantify whether two drugs together are synergistic... using the standard reference models', and the when is a concrete trigger clause — 'Use when you have measured single-drug and combination effects (inhibition/viability) and need a synergy score'. It additionally states what it explains (model choice, data needs, score reading), leaving nothing implied. Not 4 because the 'when' is already fully explicit and specific, not merely adequate.

Trigger Term Quality

Natural user phrases are present: 'drug-combination', 'synergistic, additive, or antagonistic', 'single-drug and combination effects (inhibition/viability)', 'synergy score'. It stops short of anchor 5's comprehensive synonym coverage — terms like 'drug interaction', 'combination therapy', 'isobologram', or 'combination index' as a standalone trigger phrase are absent, so a few natural variants a user might say would not literally match.

Distinctiveness Conflict Risk

A clear niche (computing reference-model synergy scores from measured effect data) with an explicit exclusion boundary — 'NOT for looking up pre-computed synergy in a database (use the SYNERGxDB tool / cell-line-profiling skill)' — actively routes away from the nearest competing skill. Trigger phrases (measured effects, synergy score) are domain-specific and would not fire for unrelated skills. Not 4 because the negative boundary plus the routing pointer reduce conflict risk to near zero.