Skip to main content

Skills

A skill is a reusable instruction sheet that teaches an agent how to do something.

If the agent's prompt is its job description, a skill is a how-to guide in its desk drawer: "this is how we write a permit decision", "this is how we triage an inbox". Write it once, install it on any agent that needs it, and share it with other teams — or with the world.

Why not just put it in the prompt?​

You could. But a prompt belongs to one agent, and it gets unwieldy fast. A skill is separate because it is:

  • Reusable — install the same skill on five agents.
  • Shareable — publish it to GitHub, or install one someone else wrote.
  • Reviewable — it has a lifecycle, so a skill from outside gets checked before an agent is allowed to use it.
  • Improvable — refine the skill and every agent using it improves at once.

Rule of thumb: if it describes how to do a task well, it is a skill. If it describes who this agent is, it belongs in the prompt.

What a skill looks like​

Skills use the open agentskills.io format, so they are not locked to Hermiq. A skill is a Markdown document — a SKILL.md — with two parts:

  • Frontmatter — a short YAML block at the top: the skill's name, a description, and metadata.
  • Body — the actual instructions, in plain Markdown.

A skill can also carry auxiliary files — extra documents alongside the main one, for templates or examples.

Because it is just Markdown, you can read a skill before you trust it. There is no hidden code.

Writing a skill​

Hermiq gives you two ways, and they suit different people:

Write or paste it directly. Open Skills → Add Skill and write the SKILL.md in the built-in Markdown editor, or paste a whole agentskills.io package and let Hermiq split it into frontmatter and body. Good when you already know what you want to say.

Talk it through with an agent. Hermiq ships a skill-creator skill. Install it on any agent, then simply chat: describe the task, and the agent interviews you and drafts the skill. When you are happy, Save as skill on its reply opens the editor pre-filled, so you review and adjust before saving. Good when you know the job but not how to write it down.

Skills created from a chat land quarantined — see below.

The review gate​

A skill is instructions an AI will follow, so where it came from matters.

  • Skills you write yourself are active immediately.
  • Skills that come from somewhere else — installed from GitHub, or produced by an agent in a conversation — land quarantined. An agent will not use them until a person clicks Approve.

This is deliberate. A quarantined skill is one nobody has vouched for yet, and "an AI wrote it" is not vouching. The review is a human reading the Markdown and deciding it says what it should.

Sharing skills​

Skills travel through the Store:

  • Find — browse skills published on GitHub and install them (quarantined, pending your review).
  • Publish — push a skill you wrote to a GitHub repository so other teams can install it.

Because the format is the open agentskills.io standard, a skill you write here is not trapped here.

Where to find it​

Skills in the Hermiq navigation lists every skill, with actions to export, publish, approve and install onto an agent. Store is where you find skills published by others.

  • Agents — skills are installed onto an agent.
  • Context — the other kind of text you give an agent, and the one skills get confused with.
  • Concepts overview — the skill-vs-context distinction, in one rule of thumb.