AI Agent Skills: The Unit That Replaced Prompts

An AI agent skill is a folder of instructions an agent loads on demand, holding the procedure for one job, so the capability lives in version control rather than in a prompt somebody retypes from memory.

We run 27 of them at ChatFuse. They cover briefing, code review, deployment preflight, closing the books, drafting outbound writing, and the weekly sprint reset. None of that lives in anyone's head or in a saved prompt any more.

The reason to care is not tidiness. A prompt cannot be reviewed by anyone except the person about to paste it, and it cannot fail loudly. A skill can do both.

What is an AI agent skill?

An AI agent skill is a directory containing a SKILL.md file, a YAML frontmatter block naming the skill and describing when to use it, and optionally the scripts that skill needs. The agent reads the description, decides the skill applies, and loads the full instructions only then.

That last part matters more than it sounds. Loading 27 full procedures into every conversation would bury the actual work. Loading one line per skill and the full text only on demand is what makes a large library usable at all.

How is a skill different from a good prompt?

A prompt is a message. A skill is a file, which means it has a history, an owner, a diff when it changes, and somewhere for a fix to live permanently. When a prompt produces a bad result you fix it in that one conversation. When a skill produces a bad result you fix the file and every future run inherits the fix.

The practical test is what happens when the procedure is wrong. Our outbound writing rules changed 4 times before they were right. As a prompt that would have meant 4 rounds of me remembering to paste the newest version. As a skill it meant 4 commits.

There is a quieter benefit. Because the file is the procedure, somebody else can read it and disagree with it. A prompt buried in one person's history cannot be argued with, so it never improves except by accident. Half the ChatFuse skills got materially better because a step in them looked wrong when written down, and nobody would have seen the step at all if it had stayed in a chat window.

What goes inside a skill?

Instructions written for an agent rather than a person, plus whatever the job actually needs. Ours vary from a page of rules to a directory with scripts, a stylometric profile, and a corpus of real examples.

What happens when a skill is written wrong?

It can disappear without telling anyone, which is the failure we hit. We run this repo from 2 different agent harnesses, and they disagree about a missing frontmatter block. One tolerates it and infers a name from the folder. The other refuses to load the skill at all and writes the error to a log nobody reads.

The result is a capability that works perfectly in one place and silently does not exist in the other. Nothing errors in the session. The agent simply never offers the skill, and you assume it chose not to. We now treat the frontmatter block as mandatory and the skill name has to match its directory.

Do skills replace prompts entirely?

No, and treating everything as a skill is its own mistake. A one off question is a prompt. Anything you have now done thoughtfully 3 times, and would be annoyed to get wrong on the fourth, is a skill.

The signal we use is repetition plus consequence. Asking a model to summarise a page is a prompt forever. Drafting anything that leaves the building, deploying, or closing a client session are skills, because the cost of a slightly different procedure each time is real.

The cost is worth naming precisely. A deploy done slightly differently each time is how a step gets skipped, and the skipped step is never the obvious one. Writing it down does not make the agent smarter. It removes the chance that this run quietly differs from the last one in a way nobody notices until production.

How do skills fit with running several models?

They sit above the model, which is the point. A ChatFuse skill describes a job rather than a model, so the same procedure runs whichever model is answering, and the ChatFuse Orchestrator still routes each step to whatever suits it across more than 100 models from OpenAI, Anthropic, Google and Meta.

That separation is what lets a skill outlive the model it was written against. Models get retired on the provider's schedule, which we covered in AI model deprecation. A procedure written as a skill survives that; a prompt tuned to one model's quirks does not.

Frequently asked questions

How many AI agent skills should a team have?

Start with 1 and add a skill the third time you do something the same way. We run 27 at ChatFuse and they accumulated over months rather than being designed up front. A library nobody uses is worse than no library, because it still has to be maintained.

What should a skill description say?

When to use the skill, in the words someone would actually use to ask for it. The description is the only part of a skill that stays in context permanently, so it is doing retrieval work rather than documentation work. A description that explains what the skill is, rather than when it applies, never gets loaded.

Can AI agent skills be shared between different AI tools?

Sometimes, but check the loader rules first. We keep 1 canonical copy and symlink the vendor specific paths to it, so both harnesses read the same file. The catch is that different tools disagree about malformed skills, and one of ours drops a skill with no visible error rather than guessing.

Are skills the same as tools or functions?

No. A tool is something the model can call, like a search or a file write. A skill is a procedure telling the model how to do a job, including which tools to use and in what order. A skill usually uses several tools.

Where do skills live in ChatFuse?

You build the equivalent in Personas, writing the job description and the boundaries into the persona so the role persists instead of living in a prompt you retype. Our AI org chart covers how those roles hand work to each other.

Skills are the least glamorous piece of running AI seriously and the one that compounds. Every fix goes somewhere permanent rather than into a conversation that ends.

Start free with ChatFuse and build the first one, or see what is included on the pricing page.