Registry Skills & Bare Skills

A skill is a markdown playbook the agent loads on demand. There are two ways to source one, and the shape of the entry tells the generator which you mean: pull a published skill from the registry (source + version), or scaffold a blank one locally (bare: true) and fill in the body yourself. This agent uses both.

apiVersion: adl.inference-gateway.com/v1
kind: Agent
metadata:
  name: research-agent
  description: Research agent mixing registry-sourced skills with a proprietary bare skill
  version: "0.3.0"
  tags:
    - research
spec:
  capabilities:
    streaming: true
    pushNotifications: false
    stateTransitionHistory: false

  agent:
    provider: anthropic
    model: claude-sonnet-4-7
    systemPrompt: |
      You are a research assistant. Consult the available skills before
      answering questions about prompting or internal review policy.

  skills:
    # Pulled from the registry - downloaded at build time, pinned by version.
    - id: prompt-engineering
      source: registry/anthropic/prompt-engineering
      version: "1.2.0"
      tags:
        - prompting
    - id: citations
      source: registry/anthropic/citations
      version: "0.4.1"
      tags:
        - research
        - citations
    # Scaffolded blank - proprietary, body filled in locally.
    - id: internal-review-policy
      bare: true
      name: internal-review-policy
      description: How to run an internal red-team review before publishing research externally.
      license: Proprietary
      tags:
        - policy
        - review

  server:
    port: 8080

  language:
    go:
      module: github.com/example/research-agent
      version: "1.26"

Highlights

• Registry-sourced (source + version). prompt-engineering and citations are published skills. The generator downloads the matching SKILL.md at build time; pinning a version keeps builds reproducible. You don't write name/description/license - they come with the published skill.
• Bare (bare: true). internal-review-policy is proprietary and one-off, so it's scaffolded blank. Here you do supply the metadata (name, description, license); the generator emits an empty SKILL.md shell with that frontmatter and you write the body.
• Licences travel with the playbook. The Proprietary licence on the bare skill ends up in its SKILL.md frontmatter. See License identifiers for the accepted set.
• Lazy loading. Only each skill's name + description is in the startup prompt; the body is read at runtime when the model reaches for it - so a manifest can declare many skills cheaply. (adl-cli auto-wires the read built-in tool whenever spec.skills[] is non-empty.)