What is an agent skill?

What is in a SKILL.md?

Three parts. At the top sits a short block of frontmatter with a name and a one-line description. Under that come the instructions, written in plain Markdown: the steps, the questions to ask, the order to ask them in, the rules that never bend. Next to the file can sit a references folder with the longer material the instructions point to, such as a checklist, a style guide, or a sample of a finished document.

The name is what shows up in your AI app. The description is the important line, and most people write it badly. It's not a summary. It's a trigger. It tells the model what a person would ask and what the skill does about it. A good one reads like "Use when a carrier estimate comes back and you want to know what to challenge. Reads the pasted estimate, lists items to question, drafts the counter to review." A bad one reads like "Settlement helper."

The body stays short on purpose. Anything long, like a 40-item inspection checklist, goes in references, and the instructions say when to open it. The model reads the body every time the skill runs. It reads a reference only when the body sends it there.

How does the model decide to load a skill?

Your AI app keeps the name and description of every skill it has in front of the model at all times. The full instructions stay out of the way until they're needed. When you type a request, the model compares it against those descriptions. If one fits, it pulls in that skill's instructions and follows them. If none fits, it answers the way it would have anyway.

This is why the description does so much work. "Turn my inspection notes into a report" fires a skill whose description mentions inspection notes and reports. It won't fire one whose description says "Reports v2." You can force it by naming the skill in your request, and that's a fine habit while you learn what a pack can do. After a week you stop bothering, because the descriptions catch what you mean.

One more detail that matters when you subscribe rather than download. When a skill is delivered over a connector, the app asks the seller's server for the instructions at the moment it needs them. So the version you get is whatever the seller published most recently, and there's no stale copy sitting on your machine.

Why does an expert's skill beat a prompt you wrote yourself?

A prompt is what you remember to say. A skill is what someone who does the job every day has learned to say, plus what they've learned not to say. The difference shows up in the edges.

Take a carrier letter. Your prompt says "write a supplement request." The public adjuster's skill says: draft only two kinds of letter, ask who's signing before you write a word, refuse if the signer isn't licensed or the policyholder, cite only the policy language the user pasted, never invent a dollar figure, and hand back a list of what still needs to be checked. Every one of those lines came from a letter that went wrong once. You'd get there too, after enough wrong letters. The skill skips you past them.

There's a second difference. A prompt you wrote in March is the same prompt in September. The model changed twice in between, the tools it can call changed, and the way it reads a long instruction changed. A skill somebody maintains gets fixed when that happens. Yours doesn't, and you find out when the output goes soft.

What does "works in your AI app as-is" mean?

Some skills are pure instructions. Paste what you have, answer the questions, get the result. Those run anywhere the model runs: the web app, the desktop app, the phone app, a coding editor. Nothing to install beyond the connection itself.

Other skills carry a scripts folder. The instructions say "run this Python file on the spreadsheet" or "call this shell command." Those need a machine that can run code, which means a developer's environment. Open one in a chat app and you watch the model try to run a script it can't reach, then apologize. That's not a skill problem so much as a delivery problem, and it's why we mark every skill with where it runs. On ProvenSkills, a pack has to contain at least one skill that runs anywhere, and skills that need scripts are never served to a chat app. If a pack page says a skill works as-is, that's what it means: no scripts, no local setup, it runs where you already are.

There's a middle ground worth knowing about. A skill can read and write your real systems without any scripts, as long as your AI app already has a connection to that system. The Claims Desk pack is a good example. Its skills tell the model how to use an AdjustCRM connection you've already added, so the skill stays plain text and the work still lands in your CRM.

Why subscribe to skills instead of downloading a zip?

You can buy a zip of skills for $9 to $15 in a few places. It works on the day you buy it. Here's what happens after.

The seller finds a bug, or the model changes, or a carrier changes its form. So they fix the skill. Your zip doesn't know. If the seller emails you a new zip, you unpack it, find the folder, replace the files, and hope you didn't skip a version. Most people don't do that. They keep running the old one until it stops being useful, then they stop using it.

A subscription flips that. Your AI app asks for the current version each time. The seller publishes a fix on Tuesday and you're running it on Tuesday, with nothing to do. When you cancel, access runs to the end of the period you paid for and then stops. You keep everything the skill already produced. That's the deal: the seller keeps the skill current, you pay while it's earning its keep, and neither side has to manage files.

Free packs work the same way. The free packs on ProvenSkills are delivered through the same connection and get the same updates. The only difference is the price.

How does a pack reach my AI app?

A pack is a named set of skills from one seller, sold per organization with a tier you pick. Once you have one, you connect your AI app to the seller once. On Claude that's one link that opens the add-connector dialog with everything filled in, then an email code to sign in. On Cursor it's a similar link into its server settings. Step by step: add skills to Claude or add skills to Cursor.

From then on every skill in the pack shows up by name in your app. Ask for the job in your own words and the right one fires. The connection can only send skill text one way, to you. It can't read your chats, your files, or your other connections. If a skill uses your CRM, it does that through your own CRM connection, under your own login.

Where to start

Browse the skill packs to see what's published, or start with a free pack and connect it in about two minutes. If you're a public adjuster, the Claims Desk Field Kit is free and covers intake, inspection notes, carrier letters, deadline research, and settlement counters. Read how it works for public adjusters for a worked example of each.

And if you're the expert with the skills, sell them here. You bring the file. We handle the rest.