Skills

Skills package reusable instructions, references, and scripts so an agent can load procedural guidance only when a task needs it.

Explore skills

Page	Learn how to
Directory structure	Organize a local skill package.
Write SKILL.md	Define metadata and focused operating instructions.
References and scripts	Add supporting files and deterministic helpers safely.
Load skills	Load one or more trusted skill sources.
Skill tools	Understand the generated tools available to the model.
Validation	Catch invalid packages before serving requests.
Skills and retrieval	Choose procedural guidance or factual search.

Progressive disclosure

Skill names and descriptions
          ↓
Load relevant SKILL.md instructions
          ↓
Read one reference or run one script if needed

The agent initially sees a compact catalog. It can load deeper content through generated skill tools instead of placing every instruction and reference in every prompt.

A minimal skill

skills/
  release-notes/
    SKILL.md
    references/
      style-guide.md
    scripts/
      collect-changes.sh

---
name: release-notes
description: Draft release notes from product changes.
---
# Release Notes

Read `references/style-guide.md` before drafting.
Use `scripts/collect-changes.sh` when change data is not supplied.

Load the directory and attach the resulting SkillSet:

import { AgentBuilder } from '@anvia/core'
import { loadSkills, skill } from '@anvia/core/skills'

const productSkills = await loadSkills(skill.local('skills'))

const agent = new AgentBuilder('release-assistant', model)
  .skills(productSkills)
  .defaultMaxTurns(4)
  .build()

Choose skills for procedures

Skills answer “how should this task be performed?” Retrieval answers “which facts are relevant?” Use skills for workflows, rubrics, and operating procedures; use Dynamic context for a large or changing factual corpus.