Living Skills

Body on disk, metabolism in the Brain

A skill has two halves that live in two different places on purpose. The procedure is a real SKILL.md file, materialized to disk so a CLI harness (Claude Code, for instance) loads it through its own native progressive-disclosure loader — Agentis injects nothing into the harness's prompt to make this work. Its discoverable description, confidence score, and links to lessons and examples live as a skill atom on the Brain's dedicated skill_library plane. SkillService is the one place that reads and writes that atom and projects it back out to a SKILL.md string — so the two halves never drift out of sync.

Cheap to search, complete when retrieved

The atom's summary is the cheap, searchable description used to find a relevant skill quickly; the full procedure — the entire SKILL.md body — travels with the atom in a separate details column, so retrieval stays fast without needing a second round-trip to fetch the actual content once a skill is chosen.

Upsert by slug, scoped, never duplicated

A skill is identified by a slug within its scope — workspace-global, or bound to a specific agent, app, or workflow. Writing a skill with a name that already resolves to an existing slug in that scope updates the existing atom in place rather than creating a second one, which is exactly what makes re-importing an agent's SKILL.md files safe to run repeatedly.

Continue