Agent skills: how to teach an AI agent a fixed workflow

Agent skills are fixed workflows you give an AI agent. A skill is a folder with an instruction file and sometimes scripts or examples. The agent opens it only when a task calls for it. That way, you don’t have to explain again and again how your team writes a quote, drafts a report, or runs a check.

You see the word skills everywhere around AI agents. But what exactly is a skill, and how is it different from a prompt, a tool, or an MCP server? In this article, you’ll learn what a skill looks like and how Hermes works with skills. You’ll also learn how to manage skills across a team and where the risk lies with skills made by others.

A row of dark slabs on a black plinth, with one cobalt-blue slab rising up and glowing softly
A skill waits until a task calls for it. Only then does the agent open the workflow.

What are agent skills?

A skill is a folder with at least one file: SKILL.md. At the top of that file are a name and a short description. Below that is the workflow itself: the steps in the right order, the pitfalls, and how to check that it worked. The same folder can also hold scripts, reference material, and templates. That is how the Agent Skills open standard describes it.

A small skill for the quote example could look like this:

---
name: write-quote
description: Use when writing a quote for a new customer.
---
1. Read the customer's request and list what they asked for.
2. Fill in the quote template in templates/quote.md.
3. Check every price against the current price list.
4. Send the draft to a colleague before it goes out.

Anthropic developed the format and released it as an open standard in December 2025. According to the specification, only the name and the description are required. The name can be up to 64 characters long, the description up to 1,024. So a skill is plain text that you can read, edit, and keep in version control. Many agents and coding tools work with this format, including Hermes. A skill you write once can be used in several tools.

How does an agent know which skill it needs?

An agent does not read all skills at once. It works in three steps. At startup, the agent knows only the name and the description of each skill. When a task matches a description, the agent reads the full instructions. During the work, it then opens the scripts or reference material that come with the skill, but only when needed.

That is why the description matters so much. If it vaguely says "helps with documents," the agent can’t tell when the skill fits. A good description says when to use the skill, for example "use when writing a quote for a new customer." And because only those short descriptions are always loaded, an agent can keep many skills ready without making every conversation heavy. The specification puts that first step at about 100 tokens per skill and recommends keeping the full SKILL.md under 5,000 tokens. Anthropic’s documentation gives the same figures.

Skill, prompt, tool, or MCP server: what is the difference?

A prompt is a one-off instruction, and a skill is a fixed workflow. A tool is a function the agent runs, an MCP server is a connection to another system, and memory holds short facts. They work together, but each does something different. An MCP server gives the agent access to a system, such as your CRM. A skill describes how the agent handles a task in that system. If the agent has no access to a system, a skill will not fix that: you need a tool or an MCP server first. You can read more about these connections in what is MCP.

TermWhat it isHow long it lastsExample

Prompt

An instruction you type in the conversation

Only in that conversation

"Summarize this email in three sentences"

Skill

A fixed workflow in a folder with SKILL.md

Stays in place; the agent opens the skill when a task calls for it

How your team writes a quote, step by step

Tool

A function the agent can run itself

As long as the tool is enabled

Fetching a web page or reading a file

MCP server

A connection that offers tools and data from another system through an open protocol

As long as the connection is active

Looking up customer data in your CRM

Memory

Short facts the agent always brings along

In every conversation

Your name, your role, and your preferences

How does Hermes work with skills?

Hermes Agent follows the same open standard. According to the skills documentation, all skills live in one folder per profile. Every skill is also a command. Type a slash and the name of the skill, in the terminal or in a chat channel, and the agent opens that workflow right away.

Hermes also writes skills on its own. When the agent has worked out a multi-step task, fixed an error, or you correct its approach, it saves that as a skill. Next time, the agent does not start over. With the /learn command, you can turn a manual, a web page, or a workflow you just showed the agent into a skill. Hermes keeps small facts that always apply in its memory. Longer workflows go into a skill.

By default, Hermes writes those skills without asking. If you want to see every change first, turn on approval for skills. A new or changed skill then waits until you approve it. For a business, that is often the better choice. You can read more about what Hermes does in Hermes Agent explained. How Voltti sets up the agent for businesses is on the page about the Hermes AI agent.

How do you manage skills across a team?

Our advice for teams: put your team’s agent skills, plug-ins, and MCP servers in a GitHub repo owned by your company. Install them on every Hermes setup from there. A skill that lives only on one colleague’s laptop does not help the rest of the team. And if everyone keeps their own version, the agent does the same job three different ways. With one shared repo, the team works on one workflow, and you can see who changed what.

Hermes supports this in a few ways. You can add a GitHub repo with skills as a source and install skills from it. You can also have Hermes read an extra folder, such as a copy of that repo on the computer where the agent runs. And you can have the skills the agent writes itself land in a folder under version control. Here is how to go about it:

  • Start small, with one skill for a task that comes back every week, such as a report or a quote.

  • Write a description that says when the skill fits, and keep the workflow short and concrete.

  • Have a colleague review every change to a skill, just as you would with code.

  • Turn on approval for skills so the agent can’t change the team’s workflow on its own.

  • Be careful with a shared folder. According to the documentation, Hermes can edit skills in a shared folder if it has write access there. If you don’t want that, make the folder read-only.

If your business has several departments, each department can get its own profile with its own skills. You can read how that works in Hermes for teams.

Can you use skills made by others?

Yes, and that can be useful: there are public collections with thousands of agent skills. But a skill is more than text. It can contain scripts that the agent runs. It can also contain instructions that make the agent do things you don’t want. A January 2026 study (a preprint) analyzed 31,132 skills from two large public marketplaces and found at least one vulnerability in 26.1% of them.

Hermes runs a skill from an outside source through a security scan first. The scan looks for data leaks, hidden instructions, and commands that cause damage, among other things. If the scan marks a skill as dangerous, Hermes won’t install it, even if you try to force it. Still, a scan does not replace your own review. Read a skill before you install it, and pay special attention to the scripts. Anthropic gives the same advice: "We recommend installing skills only from trusted sources."

You’ll find more measures in our checklist for AI and business data. For how Hermes compares to another open-source agent with its own skills, read Hermes vs OpenClaw.

How does Voltti use skills itself?

Our own Hermes setup at Voltti uses skills for work that comes back. A skill for content planning, for example, sets which Search Console data the agent pulls and how it turns that data into a content calendar. Each environment has its own settings and its own skills. That way, the agent does not use the workflow of one job for another. The agent keeps our knowledge base with the official llm-wiki skill, which you install separately in Hermes. That knowledge base has one note per topic with its sources, plus an index and a log. During a security audit, the agent follows Voltti’s own skills, with fixed work instructions for each type of check. You can read how that works in how Voltti runs a security audit.

A skill does not make an agent smarter than the language model behind it. What a skill does is make the agent handle the same task the same way every time, with the lessons from last time included. To see how an agent also runs this kind of work on a schedule and with helper agents, read agentic workflow.

Where do you start with agent skills?

Pick one task your colleagues do the same way every week, and write down how it goes. That is your first skill. Let the agent run the task a few times, compare the result with your team’s work, and improve the skill. In a pilot, you test this with the access the task needs, and no more. A pilot usually takes 2 to 4 weeks, from the first conversation to the evaluation. You can read how that works in our approach.

Portrait of Seppe Gadeyne

Seppe Gadeyne is a developer and the owner of Voltti. He sets up AI agents for companies in Belgium and the Netherlands.

Updated on

Frequently asked questions

01What is the difference between a skill and a prompt?

A prompt is an instruction for one conversation. A skill is a fixed workflow that stays in place. The agent opens the skill on its own when a task calls for it, or you call it up with a command. That way, you don’t have to type the same explanation over and over.

02What is the difference between a skill and an MCP server?

An MCP server gives the agent access to another system, such as your CRM or your calendar. A skill describes how the agent handles a task, including with that access. You often use them together: the connection fetches the data, and the skill says what the agent does with it.

03Can I use skills made by others?

Yes, but read them first. A skill can contain scripts and instructions you don’t want. A January 2026 study found at least one vulnerability in 26.1% of the skills it analyzed from public marketplaces. Hermes scans skills from outside sources before installing them, but that scan does not replace your own review.

04Where do you store skills for a team?

We recommend a GitHub repo owned by your company for your team’s skills, plug-ins, and MCP servers. From there, you put them on every Hermes setup. Hermes can add that repo as a source, or read a copy of it as an extra skills folder. That way, everyone works with the same workflow, and a colleague reviews each change first.

05Do agent skills work outside Hermes?

Yes. Agent Skills is an open standard, and a skill that follows the specification works in any agent or coding tool that supports the format. The list of supported tools includes Claude Code, Codex, Cursor, GitHub Copilot, and Gemini CLI, and Hermes Agent is on it too. So you can reuse a skill you wrote for Hermes in those tools.

Which workflow do you want your agent to learn?

Discuss your pilot