Agent Skills Explained: When Claude, Cursor, and Codex Actually Use Them

A prompt is something you remember to use

A skill is something the agent remembers to use for you. The difference sounds small until you have pasted the same prompt five times in one afternoon and tweaked it slightly each time.

A skill is a folder with a SKILL.md file inside. The file teaches Claude, Cursor, Codex, or any compatible agent how you want a specific job done once, then auto-loads when the task comes up. The source calls this lazy loading: only the one-line description sits in context, the instructions load when triggered, and reference files load only when referenced. The context cost of twenty installed skills is near zero for the nineteen that are not in use.

How a skill is actually built

The folder name is lowercase with no spaces. Inside it, SKILL.md has two parts. A short header tells the agent when to use the skill. The instructions tell it how. That is the whole format.

The same folder runs in the Claude app, Claude Code, and the API. As of late 2025, Skills became an open standard read by Copilot, Cursor, Codex, and more than forty other tools. The portability matters because you are not betting on one client. You are betting on a file format that follows you across agents. A prompt locked inside one chat window dies when the interface changes. A folder of SKILL.md files moves with you.

The heavy material can live in the same folder: brand guides, style sheets, example outputs, reference documents. A skill can know a fifty-page brand book without dragging it into every conversation. The description is always in context. The instructions load only when triggered. The reference files load only when referenced. That is why twenty installed skills do not blow up your context window.

The trigger line is the whole product

The description is the trigger. The agent decides whether to use the skill based on that line alone. If it does not match the words you actually type, the skill never fires.

Write descriptions as Use when I followed by the exact phrases you type. Vague descriptions make the skill invisible. Overly broad descriptions make it fire at the wrong time. The skill that never triggers is worse than no skill at all.

Match your real language, not your ideal language. If you usually say check my draft for facts, the description should say Use when I ask you to check facts in my draft, not Use when I need rigorous fact verification. The agent cannot read your mind. It can only read the line.

Four skills to steal

These four cover the most common pain points that make people paste prompts repeatedly. Each is a complete SKILL.md file that you can drop into a folder, upload, and edit to fit your work.

• Your Voice On Tap. Stop re-explaining how you write. The skill captures tone, sentence length, jargon rules, and what to avoid.
• The Fact-Checker. Catches invented numbers before they go out. It runs when you ask for a review of claims, sources, or stats.
• The Decision Filter. Kills bad ideas before you spend a day on them. It asks the questions you always wish you had asked.
• The Session Saver. Stops long chats from going dumb. It summarizes context, restates goals, and flags when the thread has drifted.

Each takes about five minutes to set up. None of them require code.

When skills break

Skills fail in predictable ways, and most failures come from the description, not the instructions.

• Never fires. The description does not match your real language. Rewrite it as Use when I with the exact phrases you type.
• Fires too often. The description is too broad. Narrow the job and remove vague triggers.
• Ignores its own rules. The instructions are too long, contradictory, or trying to do too many jobs. One skill, one job. Split the rest into separate folders.
• Stale. The instructions no longer match how you work. Edit SKILL.md and re-upload.

Skills are living files, not one-time setups. They degrade silently if you do not update them. Check your active skills monthly and prune the ones that no longer match how you work.

What to skill first

Rule one: if you have pasted a prompt more than three times, it is a skill. Rule two: if you have explained the same preference more than twice, it is a skill.

Start with the one that annoys you most. The prompt you keep digging out of your notes app. The explanation you keep retyping at the top of every session. That is your first skill. The best first skill is rarely the most impressive one. It is the one you are already tired of typing. The people getting leverage out of agents are not writing better prompts in the moment. They built their best thinking into skills once, so every session starts with the model already knowing how they work.

Which tasks should become skills

The shift

You have that one prompt. The good one. The one you keep in a notes app and paste into Claude five times a day, tweaking it slightly every time, wondering why you are still doing this by hand.

That prompt is now a skill. You write it once, upload the folder, and the agent loads it automatically every time the right task comes up. No pasting. No tweaking. No notes app.

The shift that matters: you stop being a prompt operator and become a system builder. Your best thinking lives in the filesystem, not in your head. The agent reads it every time. You never paste the same prompt twice.

Where to start

Start in the agent you already use. Claude Code, the Claude app, Cursor, and Copilot all read the same SKILL.md format. If you are already using Claude Code, see the guide on CLAUDE.md in practice to keep your instruction files from growing out of control. For a broader look at making coding agents reliable, read getting more out of a coding agent.

Tool pages worth bookmarking: Claude Code, Cursor, OpenAI Codex, and GitHub Copilot.

The honest summary

Skills are not a feature you turn on. They are a habit you build. The reward is not one great prompt. It is a library of your own judgment that loads automatically in every session.

Build the skill before you need it again. Start with the prompt you paste most often. Make the description match your real words. Keep one job per skill. Update it when your workflow changes. The agents that feel smartest are usually the ones whose owners did this work first.