Publishing skills

A skill is a self-describing capability an LLM client can load. ARP §6.1 adopts the agentskills.io open spec and the Anthropic Claude Skills frontmatter format verbatim — your skill is portable to all four channels at once.

SKILL.md format

Every skill ships as a single SKILL.md file with YAML frontmatter and a Markdown body:

---
name: my-skill
description: One-sentence summary of what this skill does. Trigger on phrases like "do X", "search for Y".
version: 1.0.0
license: MIT
allowed_tools: Read, Glob, Bash
auto_invocable: false        # default — opt-in to auto-load (safer)
metadata:
  owner: my-team
  tags: [observability, eval]
oasf_skill_ids: [1106]       # optional AGNTCY OASF taxonomy ref
---

# My Skill

What this skill does. When to invoke. Steps to follow. Hard rules.
The body is freeform Markdown — Claude Code reads it verbatim.

Required vs optional

FieldStatusNotes
nameREQUIREDLowercase kebab-case. Must match directory name.
descriptionREQUIREDTells the LLM when to invoke. List trigger phrases.
versionoptionalSemver. Bump on breaking changes.
licenseoptionalSPDX identifier (MIT, Apache-2.0, CC-BY-4.0...).
allowed_toolsoptionalComma-separated allowlist of tools the skill needs.
auto_invocableoptionalDefault false. Opt-in to LLM auto-loading. (Polarity-flipped from Anthropic's disable-model-invocation.)
metadataoptionalFree-form key/value bag (owner, tags, dashboard URLs).
oasf_skill_idsoptionalAGNTCY OASF taxonomy IDs (LF AI & Data alignment).

Where to publish

One SKILL.md, four distribution channels:

  1. Anthropic Skills marketplace — PR to anthropics/skills. Reach: 125K+ Claude Code users.
  2. agentskills.io — web upload UI. Reach: 30+ runtime adopters.
  3. AGNTCY Hosted Agent Directory — OASF descriptor at /.well-known/oasf-descriptor.json. Reach: LF AI & Data ecosystem. (AgentMinds publishes one.)
  4. Your own repo — anyone with a clone can use it. Reach: viral / org-internal.

Reference: AgentMinds' own skills

← Getting started|Conformance levels →