What Are Agent Skills?

Agent skills are folders of instructions, built around a file called SKILL.md, that tell an AI how to do one specific job. The AI reads a short description of each skill and loads the full instructions only when a task matches. The format is an open standard, so one skill can work in many AI tools with little or no change.

Picture a capable new starter. On day one they can write, research and analyse, but they don't know how your company writes a proposal or reports to a client. You'd give them a job card: the steps, the format, the checks and where to find the templates. A skill is that job card, written for an AI.

A skill isn't a one-off request you paste into a chat. It's saved, it has a name and a description so the AI knows when to use it, and it runs the same method every time. It also isn't the AI's access to your systems. A skill can say "use the client notes", but the AI still needs its own permission to open them. Skills vs MCP explains the difference.

A typical business skill contains:

• A name and a short description of when to use it.
• The steps for the job, including what to do when information is missing.
• The output format: the sections, the length and who will read it.
• Checks to run before the work is finished.
• Optional supporting files, such as a template, a worked example or a script.

The Agent Skills format was first developed by Anthropic and released as an open standard that other products have since adopted (Agent Skills overview). At its simplest, a skill is a folder containing a SKILL.md file. The file opens with a short block of details about the skill, followed by the instructions in plain Markdown.

The specification sets two required fields (Agent Skills specification):

• name: up to 64 characters, using lowercase letters, numbers and hyphens, and matching the folder name.
• description: up to 1,024 characters, saying what the skill does and when to use it.

Optional folders hold scripts, reference documents and assets such as templates. The specification suggests keeping the main file under 500 lines and moving detail into separate reference files.

AI tools load skills in stages. At the start of a session they read only each skill's name and description. When a request matches, they read the full instructions. Supporting files are opened only if the task needs them. That's why a team can keep many skills available without crowding the AI's working memory. It's also why the description matters so much: it's how the AI decides whether to pick the skill up.

A shared format doesn't mean identical behaviour. Each product decides how it loads skills, which extra fields it reads, whether scripts can run and which tools a skill can call. Test the cases that matter in every tool your team uses.

Northgate Advisory, a fictional consultancy, sends each client a weekly project update. Without a skill, each consultant asks their AI in a different way and the updates come back in different shapes. Some miss the risks. Some state a proposed decision as agreed.

A first skill for this job might look like this:

• Name: weekly-project-update
• Description: Turns this week's project notes into a client update covering progress, decisions, risks and next actions. Use when someone asks for a weekly update, status report or progress note.
• Inputs: this week's notes, last week's update and the client's name.
• Steps: compare with last week, list what moved, separate agreed decisions from proposals and give each risk an owner.
• Missing information: if an action has no owner or date, mark it "owner needed" rather than guessing.
• Format: four headed sections, under 400 words, written for the client's project lead.
• Check before finishing: every decision and date in the update must appear in the notes.

Notice what the skill leaves out. It doesn't hold the client list, the project notes or Northgate's tone of voice guide. Those change often and belong somewhere else, which is where company context and a knowledge base come in.

• Pick one job you repeat at least weekly, with a result you can judge. How to Create an AI Skill covers each step in more detail.
• Write one sentence that names the task and the result.
• List the inputs the job needs and what the AI should do if one is missing.
• Describe the output: the sections, the length and who reads it.
• Write the description last, using the words people actually type when they ask for this job.
• Save it as a folder with a SKILL.md file, or draft it in the skill builder.
• Add it to one AI tool and run the three tests below.

• A normal request. Give it a full set of notes. Check that the four sections are there and that every fact traces back to the notes.
• Incomplete input. Remove an owner or a date. The update should flag the gap, not invent a name.
• An unrelated request. Ask for something nearby but different, such as a meeting agenda. The skill shouldn't take over. If it does, narrow the description.

Keep these three examples. Run them again after every change, so you can tell whether an edit helped or made things worse.

A skill tells your AI what to do. Your knowledge base shows it how, in detail. Your company context tells it who it's for. The Contexta harness brings them together, in every AI.

Contexta holds your skills alongside your company context: who you are, the terms you use and how your work is laid out, set for the company and adjusted for each team. Skills point to the templates, SOPs and past examples in your knowledge base, which stays in your own SharePoint or Google Drive. Company, team and personal spaces give each group its own home and rules. Every change to a skill is saved in version history, so you can compare versions and restore one that worked.

Connected AI tools fetch the published skill the next time they use it, so a change reaches everyone without files being copied around. A skill you download is a separate copy and won't update itself. See How It Works for the setup and Integrations for supported AI tools.