Skills

A skill tells an agent how to approach one kind of task: which tools to use, in what order, and what a good result looks like. A skill is a SKILL.md file in the agent's workspace. The agent finds its skills with search_files and reads the one a task needs.

Skills use the Agent Skills layout. A skill made only of instructions works as it is.

Write a skill

---
name: issue-triage
description: Triage a new GitHub issue, find duplicates and suggest labels.
---
Search open and closed issues for the error message before anything else.
Name a duplicate only when the steps to reproduce match. Suggest at most two labels.

Put each skill in its own directory: skills/<name>/SKILL.md. Files a skill refers to go beside it.
description says when to use the skill. The agent reads it to choose.
A skill holds instructions only. Nothing in it runs. Work a script would do belongs in a tool or an MCP server.
A skill can't give an agent tools. The agent's tools and policy still decide what it can do.

Give an agent skills

Keep skills in a directory, such as one in git, and sync it into the agent's workspace:

wardrik agent files sync triage ./skills --to skills

Tell the agent to look for skills in its instructions:

---
tools: [files, github/*]
---
Before a task, run search_files for `^description:` in `skills/*/SKILL.md`. Read the skill that
fits the task, and follow it.

wardrik reload triage

search_files returns the description line of every skill in one call. The agent reads only the skill it needs.

Keep shared skills in one agent's workspace and mount it read-only in the others:

tools: [files]
workspace:
  shared:
    skills: {agent: agent:skills, mode: read}

wardrik agent files sync skills ./skills --delete

Tell each agent to search shared/skills/*/SKILL.md too. config.yaml sets who can mount the workspace. See Sharing workspaces.

Let an agent write skills

An agent with files can write skills of its own. Require approval for each one:

policy:
  ask: ["write_file(skills/*)", "edit_file(skills/*)"]

A skill written in a turn that read outside content is labeled outside. Reading it labels the agent's turn too, so instructions from outside content can't come back later as a trusted skill.

Skills

Write a skill

Give an agent skills

Share skills between agents

Let an agent write skills