Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents

gist

Scientific Agent Skills is K-Dense's open library of 163 agent-loadable procedures across 16 research areas, from genomics and cheminformatics to study design and scientific communication. Each skill is a directory whose versioned instruction file loads only when a task needs it, often with references and scripts. The paper reports no task-level evaluation; it measures documentation cost instead: resident descriptions of all skills take 7.1% of a 200,000-token window, and the median documented workflow fits in 23.9%, though loading every reference overflows 29 of 46 workflows.

ideas

• Working code is not defensible analysis. Field-specific procedural choices decide which test applies, which identifier namespace is authoritative, and which caveats must travel with a result.
• Skills are authored procedures, not retrieval passages. Each skill is a directory built around a human-readable instruction file the agent loads only when the task appears to call for it.
• Progressive disclosure keeps standing context small. Always-resident name-and-description rows for 163 skills cost 7.1% of a 200,000-token window; instruction bodies and references stay deferred.
• Reference loading is the overflow risk. Median documented workflows fit at 23.9% with instructions alone, but 29 of 46 exceed the window if every named reference file is loaded.
• This is a resource paper, not a task eval. The authors report no host selection rate and no controlled study that skills improve scientific outcomes.

quotes