Skills

Living skills — reusable agent procedures with a contract, versions and quality evaluations.

A living skill is a reusable agent procedure: what it does, when it applies, which tools it needs, how to verify the result. Unlike a prompt, a skill is versioned, has a contract and a suite of quality evaluations. The Skills page (/skills); details — docs/living-skills.md.

#The skill contract

The skill switch is the toggle in the bottom-right corner of the mini-card: a disabled skill is never served to agents (like an unapplied draft), its content stays untouched. Every flip is recorded in the audit log.

The manifest (SKILL.md + metadata) describes:

• the procedure — the steps the agent performs;
• preconditions — when to apply it, what must be on input;
• required tools and rights — a skill cannot get more than the agent has;
• postconditions and checks — how to verify completion;
• artifacts — reference documents (policies, instructions) the agent loads on demand via artifact_read, not from the prompt.

#How one is created

1. From a description — describe the repeating procedure in plain words; the builder drafts the skill referencing only tools that actually exist.
2. Import — a folder or archive in the agentskills.io ↗ format.
3. Manually — the manifest editor.
4. Mining — the supervisor finds repeated tool-call trigrams across runs (a deterministic gate filters out read-only sequences and back-to-back repeats), the judge issues a worth_skill verdict with a rationale; the candidate appears in the queue on the Skills page as a pre-filled draft — all that is left is to read and approve or reject it.

A new skill is a draft (0.x). It reaches production after human review. Badges in the skill list show the unfinished work: draft — drafts, "↺ N" — pending draft revisions; applying or rejecting a version is a human decision.

#Quality evaluations

Golden cases attach to a skill: input → expected answer. An evaluation run (manual or scheduled) checks answers with patterns and an LLM judge with agreement calibration. The result is visible in the skill's card and on the quality dashboard.

#Evolution

The skill lives: the agent can propose a change (the skill_propose tool) based on repeated friction. Versions accumulate; applying a new version goes through the same quality loop. The tools skill_history, skill_diff, skill_executions give the full picture: who changed what and when, where the skill ran and what it "remembered" from executions (skill_memory).

Artifact references ride the evolution. A change proposal (skill_propose) carries an artifact_refs list — the documents the skill needs as reference material. Each reference is validated against the registry at submit time: only approved artifacts enter the manifest, while hallucinated or unapproved ids are dropped and listed back in the response — the model sees exactly what survived. A server-side check at the save boundary makes it impossible for a skill to reference an unapproved document. References can also be edited by hand in the skill editor's "Artifacts" tab: search the registry, attach a document and describe when it applies.

#How the agent works with skills

In a run the agent has skill_search (find a match), skill_inspect (read the contract), skill_validate (check applicability). The skill is loaded into context on demand, not as a permanent prompt.