What Are Skills in Working with AI Agents

A skill is a working method that teaches an AI agent to perform a task consistently

Let’s start at the end: a skill is a set of instructions that repeats every time an agent is required to perform a certain type of work.

Think of a new employee joining a team. If every morning you have to re-explain who the client is, how you write here, what must never be sent, and how to review work, it’s exhausting. A skill keeps all of that in one place.

In plain terms

A skill tells the agent: this is how we work on this task.

What it solves

It prevents us from explaining the same working method from scratch every single time.

Before talking about skills, it helps to align on four concepts

A lot of confusion around AI comes from using the same words for different things. So let’s start with a simple separation.

A good skill answers six basic questions

For an AI agent to work well, it needs more than an opening sentence. It needs to understand what triggers the skill, what the goal is, what information to take into account, what the order of work is, what the output should look like, and how to verify it.

• When to activate itWhich tasks should prompt the agent to use this skill.
• What the goal isWhat result you want at the end, and who it's intended for.
• What context mattersDocuments, examples, target audience, tone, constraints, data, and prior decisions.
• What the order of work isThe steps the agent must go through before returning a result.
• What the output looks likeFormat, length, structure, level of detail, and writing style.
• How to verify itWhat the agent must check, and what must remain for human approval.

When there are clear answers to these questions, the agent receives a working framework that can be repeated.

When an AI agent does more than chat, you need to define a working method for it

In a regular chat, the work usually ends with a response. With an AI agent, the work can include opening files, summarizing, editing, reviewing, and returning a deliverable for further use.

That’s why skills matter more precisely when you move beyond simple conversation to agentic work. The more the agent does, the more important it is to define a goal, boundaries, and stopping points.

Start with one task that repeats itself

A good skill starts with a recurring task: summarizing a meeting, reviewing a document, drafting an email, analyzing feedback, or generating a weekly report.

• Choose a recurring taskSomething you do over and over, with a reasonably clear outcome.
• Write down what "good" looks likeLength, tone, structure, target audience, level of detail, and things that must not be missed.
• Add an exampleOne good example is sometimes worth more than ten general instructions.
• Define boundariesWhat the agent handles on its own, when it asks, and when it stops.
• Add a review stepA short list of questions the output must pass before it's used.

A good skill lives in a small folder with instructions, references, and helper tools

In a real project, it’s worth storing skills in an organized way. The goal is for each skill to be easy to read, easy to update, and easy to transfer between projects or team members.

This structure helps separate three types of knowledge: the skill’s fixed instructions, reference materials the agent can read as needed, and small tools that perform a technical action.

project/
└── skills/
    └── meeting-summary/
        ├── [SKILL.md](/en/blog/markdown-yaml-ai-agents/)
        ├── references/
        │   ├── tone-of-voice.md
        │   └── meeting-summary-example.md
        └── scripts/
            └── clean-transcript.js

• skills/ — The folder that centralizes all of the project’s skills. When there are multiple skills, this is where you see the agent’s full “work library.”
• meeting-summary/ — One folder per skill. The name should describe the task — for example, meeting summary, document review, or report preparation.
• SKILL.md — The skill’s central file. This is where you write the instructions the agent must read on every run: when to use the skill, what the goal is, what the input is, what the order of work is, what the output looks like, and what needs to be checked.
• references/ — A folder for reference materials. This is where you put things the agent should use as needed: good examples, tone of voice guidelines, templates, brand rules, glossaries, or domain knowledge.
• scripts/ — A folder for technical helper tools. This is where you put small pieces of code that perform a recurring action — for example, cleaning a transcript, checking a format, extracting text from a file, or converting data.

The simple rule: what the agent must always know goes into SKILL.md. What it only needs sometimes goes into references. A recurring technical action can go into scripts.

The human skills needed to work with AI effectively

It’s important to distinguish between a skill the agent uses and a skill the person needs to develop. To work well with AI, we need a few core capabilities that apply across every tool and every role.

• Understanding capabilities and limitationsWhere AI helps, where it falls short, and where human review is essential.
• Defining a taskWhat you want to receive, who it's for, and what counts as a good result.
• Providing contextDocuments, examples, data, prior decisions, target audience, and constraints.
• Requesting a clear outputFormat, length, tone, level of detail, assumptions, and criteria.
• Reviewing outputsVerifying facts, identifying gaps, checking sources, and understanding risks.
• Embedding into a processTurning one-off use into templates, procedures, and controlled automations.

This is what skills look like in real work

A good skill can be simple. It needs to be useful, clear, and return an output that can be verified.

• Meeting summaryTurns a transcript or raw notes into a summary, decisions, open action items, and owners.
• Document reviewChecks for clarity, redundancy, gaps in reasoning, inappropriate tone, and points requiring verification.
• Initial researchGathers directions, organizes sources, separates facts from assumptions, and suggests follow-up questions.
• Email draftWrites according to target audience, goal, tone, level of directness, and things to avoid.
• Weekly reportTakes raw information and returns status, blockers, decisions needed, and risks.

Notice the pattern: every example here is a small process with a starting point, raw materials, a desired result, and a review step.

Let’s build an example skill: team meeting summary

Say you have a weekly team meeting. Every week there’s a transcript, notes, or raw bullet points, and you want to receive a summary you can send to the team without editing it from scratch.

• Skill nameTeam meeting summary.
• When to use itAfter a team meeting, weekly status, project session, or sync call.
• What goes inTranscript, raw notes, meeting chat, participant list, and brief project context.
• What the agent doesIdentifies key topics, decisions, action items, open questions, and risks that came up in the conversation.
• What the output looks likeA short summary, decisions made, action items by owner, open questions, and topics for follow-up.
• How to verify itWas every decision actually said? Does every action item have an owner? Is there missing context that could cause a misunderstanding?

Once this structure is clear, you can turn it into a standing instruction for an AI agent. From that point on, you no longer need to reinvent the summary process every week.

Try building one skill around a task you already have

Start with one task you do at least once a week, and answer the following structure.

• Skill nameWhat would you call this task? For example: "Team meeting summary."
• When to use itIn what situation does the skill get activated.
• What goes inTranscript, document, spreadsheet, emails, notes, links, or data.
• What comes outSummary, table, task list, email draft, recommendations, or report.
• How to verify itThree questions you must ask before using the output.

If you managed to answer all five points, you already have the foundation for your first skill.

Before using a skill, run through seven quick checks

A good checklist helps you see whether the skill is truly clear, or whether it still leaves too much room for guesswork.

• Clear nameIs it obvious what the skill does just from reading the name?
• Clear triggerIs it clear when the skill should be activated?
• Clear inputIs it clear what materials need to be given to the agent before it starts?
• Clear outputIs it clear what should come out at the end and in what structure?
• A good exampleIs there at least one example showing the agent what a good output looks like?
• BoundariesIs it clear when the agent continues on its own, when it asks, and when it stops?
• Quality checkAre there review questions the output must pass before it's used?

What usually makes a skill work less well?

Skills work less well when the instructions are too general, when examples are missing, when boundaries aren’t clear, or when there’s no way to verify the output.

Basic questions before you get started