Skills

What it is. A folder with a SKILL.md (frontmatter name and description, then the instructions) and any files it uses. The agent reads the description, and the rest only when the skill is needed.

Where it goes. .agents/skills/<name>/SKILL.md (~/.agents/skills at user scope).

.agents/skills/review/SKILL.md

---
name: review
description: Review the change before you stop.
---
Read the diff and list problems.

What each agent gets

Agent	Gets
Every agent but Claude Code	nothing written: it loads .agents/skills where it is
Claude Code	a copy of each skill in .claude/skills

Run agent-harness-adapter install after you add or change a skill, so Claude Code’s copy follows; status exits 1 while it is out of date.

As a tool

Command

[[skills]]
name = "mytool"
description = "Check the project with mytool."
body = "Run `mytool check`; fix what it reports.\n"

[[skills]]
dir = "agent/skills/review"     # an existing skill folder, as it is

# or every skill folder in a directory: skills = { from = ".agents/skills" }

Rust

Integration::new()
    .skill(Skill::new("mytool", "Check the project with mytool.", "Run `mytool check`.\n"))

Node

{ skills: [{ name: "mytool", description: "Check the project with mytool.", body: "Run `mytool check`.\n" }] }

Python

{"skills": [{"name": "mytool", "description": "Check the project with mytool.", "body": "Run `mytool check`.\n"}]}

Limits

• Copilot, Cursor, OpenCode, Kilo Code and Devin also load .claude/skills, so with Claude Code installed they can see a skill twice. The install warns (may load the skills twice); the copies are identical.