Author a skill

One skill is one directory with a SKILL.md at its root. The file opens with a YAML manifest and continues with the instructions your skill gives an AI coding agent.

The SKILL.md

---
name: acme-education-issuer
description: Issue education credentials with the ACME wallet over OpenID4VCI.
version: 1.0.0
license: Apache-2.0
metadata:
  provider: ACME Wallets
  protocols: OpenID4VCI
targets:
  openapi: openapi/issuer.yaml
depends_on:
  schemas:
    - schemas/learner-credential.json
---

# ACME education issuer

Use this skill when the user wants to issue education credentials
with the ACME wallet. Read openapi/issuer.yaml for the API surface,
follow rulebooks/issuance.md for the policy rules, and validate
credential payloads against schemas/learner-credential.json.

Manifest fields

Field	Required	Rules
`name`	Yes	The skill’s slug and identity: lowercase letters, digits, and hyphens, 3-64 characters. Unique across the marketplace.
`description`	Yes	One or two sentences; shown in the catalog and read by agents.
`license`	Yes	An SPDX identifier such as `Apache-2.0`.
`metadata.provider`	Yes	Your organisation’s name.
`version`	Recommended	Your own version label; the marketplace also pins every submission to a commit.
`targets`, `depends_on`	Optional	Paths to the OpenAPI spec, schemas, and rulebooks the skill relies on. Every path must resolve inside the skill directory.

The instruction body

Everything after the closing --- is what the agent reads. Write it for the agent: when to use the skill, which files to read, which endpoints to call, what to validate. The automated checks require at least 80 characters of instructions - real skills carry far more.

Reference files

openapi/ - OpenAPI 3.x documents (.yaml or .json). Each file must parse and declare openapi: 3.x.
schemas/ - JSON schema files; each must be valid JSON.
rulebooks/ - policy rules in markdown, separable from code.
references/ - anything else the agent should be able to consult.

Keep a skill self-contained: every path in depends_on and targets must exist inside the skill’s own directory, because each skill is checked, reviewed, and installed as one unit.