Aiodic

Back to the table

106 Advanced Sk Skills Agents & Automation

Skills

Packaged instructions an agent loads when needed.

What it is

A skill is a packet of know-how that an AI agent keeps on the shelf and takes down only when a job calls for it. It's a written set of instructions for one kind of task, often with examples, sometimes with reference documents or small programs attached, and always with a short label saying what it's for and when to use it.

Think of a restaurant kitchen. The head chef doesn't recite every recipe to the cooks at the start of each shift. The recipes live in a binder. Each card has a title at the top ("Hollandaise: for eggs Benedict and asparagus"), and when an order comes in that needs hollandaise, the cook pulls that card, follows it, and puts it back. Nobody has to hold all two hundred recipes in their head at once, and the hollandaise comes out the same way every time, no matter who's on shift.

Or think of a staff handbook. A new employee doesn't read the whole thing on day one. They know the handbook has a page on "Handling a refund over £500", and when that situation arrives, they turn to it. Skills are those pages, written for an AI.

The idea became widely known in October 2025, when Anthropic, the company that makes Claude, launched a feature called Agent Skills. In that design, a skill is a folder. Inside is one required file, called SKILL.md, which is a plain text document in Markdown, a simple formatting style that uses symbols like # for headings. It starts with two labelled lines: a name and a description that says what the skill does and when to use it. Below that come the instructions themselves. The folder can also hold anything else the task needs:

The clever part is how little of this the AI reads up front. This is called progressive disclosure, which means revealing detail only as it's needed. At the start, the agent sees just the name and description of every skill it has, a line or two each. That's the title on the recipe card. Only when your request matches a description does the agent open the full SKILL.md. And only if those instructions point to a reference file or a script, and the task actually needs it, does the agent go further. The published standard suggests the name and description should cost around 100 tokens (the small chunks of text an AI reads, roughly three-quarters of a word each), and the main instructions should stay under about 5,000 tokens.

That matters because everything an AI is currently paying attention to has to fit in its context window, its limited working memory. If you poured every procedure your business uses into that window at the start of every conversation, you'd fill it with material that's irrelevant most of the time, and the AI would get slower, costlier and more easily confused. Skills let an agent have a hundred procedures on hand while carrying only their labels.

How skills differ from their neighbours. Skills sit close to several other ideas on this table, and the differences are the fastest way to understand them.

A system prompt is always on. It's the standing briefing loaded at the start of every conversation, like the rules posted on the kitchen wall. A skill is loaded only when relevant. Anthropic's help centre makes the same comparison with Claude's custom instructions, which apply to every conversation, while skills are task-specific and load only when relevant.

A prompt template is a saved prompt with blanks that you fill in and paste. You decide when to use it. With a skill, the agent decides. It reads your request, matches it against the descriptions, and pulls the right card itself. A template is also usually just words, while a skill can carry files and programs.

MCP and tool use give an AI access: the ability to open your calendar, search your drive or send an email. They don't tell it how your business wants those things done. A skill is know-how rather than access. Anthropic's help centre puts it neatly: MCP connections give Claude access to tools, while skills teach it how to use those tools well. A connector lets the agent reach your newsletter platform; a skill tells it your newsletter always opens with a personal note, never uses more than two links, and signs off with your name.

Skills are also one of the main tools of context engineering, the craft of deciding what information an AI sees at each moment. And they pair naturally with subagents: a helper sent off to do a focused job can pull the one skill that job needs.

Why it matters

It turns improvising into following a recipe. Left to itself, an AI will invent a fresh way to do a task every time. Sometimes that's brilliant; often it's merely different, and different is a problem when the job has to match your brand, your house style or your legal wording. A skill captures the way that works and lets the agent repeat it. The outcome becomes more predictable, which is what you want from anything you hand off.

It saves you repeating yourself. Most people who use AI heavily have a handful of explanations they type again and again: how they like their invoices, what their book blurbs sound like, which words their brand never uses. A skill is where that explanation goes to live. You write it once, carefully, and every future conversation gets the careful version instead of whatever you remembered to type that day.

It scales without clutter. Because only the descriptions load up front, you can add skills for dozens of tasks without slowing everything down. That's the practical breakthrough. Earlier approaches forced a choice between a short system prompt that knew little and a huge one that knew everything and drowned in it.

It can make fiddly jobs reliable. Some tasks are better done by a program than by an AI writing freehand. Anthropic's engineering article describes a PDF skill that includes a ready-made script to read a PDF and list every field in its form. The AI can run that script without reading the script itself or the whole PDF into its working memory, and the answer comes out the same way every time. For anything involving exact numbers, file formats or careful checks, that's a real gain.

It's becoming portable. Since December 2025 the format has been published as an open standard, and many AI tools now read it. A skill you write for one tool has a good chance of working in another. That makes a skill a small asset worth building, in the way a well-written procedures manual outlives whichever software a company happens to use.

It carries risk as well as benefit. A skill is instructions an agent will follow, and sometimes code it will run on your computer. Installing one from a stranger is closer to installing software than to reading a web page. That's covered in the examples and the mistakes below, and it's the one part of this topic every reader should take seriously.

Real-world examples

Anthropic launches Agent Skills (16 October 2025)

On 16 October 2025, Anthropic announced skills for Claude, describing them as folders containing instructions, scripts and resources that Claude can load when needed. At launch they worked across the Claude apps, Claude Code (its coding tool) and the developer platform. Anthropic shipped its own skills for reading and producing Excel spreadsheets with formulas, PowerPoint presentations, Word documents and fillable PDFs, which is how many people met skills without realising it: the ability to ask for a finished spreadsheet or slide deck, rather than text to copy into one, arrived partly as skills. The announcement named Box, Canva, Notion and Rakuten among the companies working with the feature.

The companion engineering article set out the design. It described the three levels: every skill's name and description loaded at the start, the full SKILL.md read only when the skill looks relevant, and extra files read only as needed. It used the PDF skill as the worked example, with its form-filling instructions split into a separate file so the main instructions stay lean.

Anthropic also put a warning in the announcement itself. Because skills can make Claude run code, the company advised being mindful about which skills you use and sticking to trusted sources. The engineering article went further, recommending that anyone installing a less-trusted skill read through its files first, paying particular attention to bundled code and to any instructions telling the AI to connect to outside network addresses.

Anthropic later published its example skills on GitHub, a site for sharing code. The repository notes that most are open source, but that the document skills for Word, PDF, PowerPoint and Excel are shared as reference examples rather than as open-source software, and it describes the whole collection as being for demonstration and education.

At the time of writing, Anthropic's help centre says skills are available on its Free, Pro, Max, Team and Enterprise plans, with a setting called code execution switched on. Plans and features change often, so check the current terms yourself.

From one company's feature to an open standard (December 2025)

Skills spread faster than most new AI ideas, partly because they're so simple: a folder with a text file in it. Any AI tool that can read files can, in principle, use one.

On 12 December 2025, the independent developer and writer Simon Willison reported that OpenAI had quietly added skills support to both ChatGPT and its Codex command-line coding tool, before the company had said anything about it publicly. He found a skills folder inside ChatGPT's code environment with skills covering spreadsheets, Word documents and PDFs, and experimental support for your own skills in Codex.

On 18 December 2025, Anthropic published Agent Skills as an open standard, meaning a shared specification that any company can adopt, now hosted at agentskills.io. The same day, GitHub announced that its Copilot coding assistant supported Agent Skills in several of its tools, with support in the regular version of Microsoft's VS Code code editor promised for early January. Willison's write-up the next day noted how short the specification is, readable in a few minutes, and that some parts are deliberately left loose. He also noted that OpenAI had added skills to its Codex documentation by 20 December.

At the time of writing, OpenAI's own help pages describe skills in ChatGPT and Codex as building on the open agent skills standard, using the same SKILL.md file with a required name and description, and loading only names and descriptions until a skill is needed. The agentskills.io site lists dozens of tools that read the format, including Gemini CLI, Cursor, GitHub Copilot, VS Code and Block's open-source agent goose. Support varies from tool to tool and changes quickly, so check your own tool's documentation.

Hundreds of booby-trapped skills (February 2026)

The downside of an easy format is that it's easy for criminals too. On 2 February 2026, the security news site The Hacker News reported research by Oren Yomtov of Koi Security into ClawHub, a marketplace of skills for OpenClaw, a popular open-source AI assistant that people run on their own computers. The researchers examined all 2,857 skills on ClawHub and found 341 malicious ones, 335 of which appeared to belong to a single campaign.

The malicious skills looked like useful tools: cryptocurrency wallet helpers, YouTube utilities, auto-updaters, finance tools and Google Workspace connectors, plus some with names chosen to look like legitimate listings. Their instructions included fake "prerequisites", steps that told the user to download a file or paste a command into their computer's terminal before the skill would work. Doing so installed malware. On Macs, it was a known password-stealing program called Atomic Stealer, which collects saved passwords, browser data and the secret keys people use to access AI services. According to the report, OpenClaw's creator, Peter Steinberger, responded by adding a way for signed-in users to report skills, with a skill hidden automatically after three separate reports.

The lesson reaches beyond one marketplace. A skill is a set of instructions that an AI, and sometimes you, will follow. A bad skill doesn't need to break anything clever. It only needs to be trusted.

An everyday scenario: the author's blurb skill

This is an illustrative scenario, not a documented event. Marcus writes fantasy novels and has spent months teaching his AI assistant how he likes book descriptions: open on the hero's problem rather than the world's history, one line of stakes, one hint of romance, never the phrase "epic journey", and no more than 150 words. Every time, he pastes the same paragraph of rules and his two favourite past blurbs.

One weekend he turns that paragraph into a skill. He names it "book-blurb", writes a one-line description ("Writes back-cover and store-page descriptions for Marcus's fantasy novels. Use when asked for a blurb, book description or back-cover copy."), lists the steps and rules, and adds one finished blurb as an example. He tests it on three books whose blurbs he already knows are good and finds it keeps adding a rhetorical question at the end. He adds a rule forbidding that and tests again. Now when he types "draft a store description for book four", the assistant pulls the skill without being told. For his newsletter, which he writes differently, it doesn't.

Try it yourself

You're going to write a one-page skill for a task you do often. It works in tools that support skills, and it works equally well as a plain document you paste into any chatbot, so you don't need special software. Allow about forty-five minutes.

  1. Pick one recurring task. Choose something you've explained to an AI at least three times: formatting your newsletter, writing a book blurb in your style, replying to a common customer question, turning meeting notes into action points. One task per skill.
  2. Collect what you already know. Find the instructions you usually type and one piece of finished work you're proud of. That example is often worth more than all the rules put together.
  3. Write the skill using the template below. The name is short, lowercase, with hyphens instead of spaces. The description is the most important line on the page, because it's how the AI decides whether to pull the card. Say what the skill does and when to use it, in the third person ("Writes…", not "I can help you…"), and include the words people are likely to use when they ask.
  4. Keep it short. A good skill fits on one page. Leave out anything the AI already knows, such as what a newsletter is. Put in only what's particular to you: your rules, your order of steps, your banned words.
  5. Test it as a plain document first. Open a fresh chat, paste the whole skill, then give it a real task using the first prompt below. Compare the result with your example.
  6. Test the edges. Try a request that should use the skill but is worded differently, and one that shouldn't use it at all. A skill that fires for the wrong job needs a narrower description; one that doesn't fire needs more of the words people actually use.
  7. Fix one thing at a time and re-test. When the output goes wrong, add or sharpen a single rule, then try again on the same tasks. This is a small version of evals, and it's how Anthropic's own guidance recommends developing skills.
  8. Install it if your tool supports skills. Save the finished page as a file called SKILL.md in a folder with the same name as the skill (for example, book-blurb/SKILL.md), then follow your tool's instructions for adding a skill. At the time of writing, Claude, ChatGPT and several coding tools accept this format, but where you upload it and which plans include it differ and change, so check your tool's current help pages.

A one-page skill template:

---
name: book-blurb
description: Writes back-cover and store-page descriptions for [your name]'s [genre] novels. Use when asked for a blurb, book description, back-cover copy or store listing text.
---

Steps
1. Ask for the book's title, the main character, their problem, and the tone, if not given.
2. Open with the main character and the problem they face, in one or two sentences.
3. Add one sentence on what they stand to lose.
4. End with a line that makes the reader want to find out what happens.

Rules
- 120 to 150 words.
- Present tense, third person.
- Never use: [your banned phrases, e.g. "epic journey", "in a world where"].
- No spoilers past the first quarter of the book.
- Do not end on a question.

Example
[Paste one finished blurb you are proud of here.]

A prompt for testing it as a plain document:

Below is a set of instructions called a "skill". Follow it exactly for
the task that comes after it. If the task doesn't match the skill's
description, say so instead of using it.

[paste your whole skill here]

Task: [a real request, e.g. "Write the store description for my new book.
Title: ... Main character: ... Problem: ..."]

A prompt for getting the AI to help you improve it:

Here is a skill I wrote and an output it produced that I'm not happy with.
Suggest the smallest change to the skill's instructions that would fix the
problem, and rewrite the description so it would be picked for the right
requests and not for others. Do not make the skill longer than one page.

Skill: [paste]
Output: [paste]
What's wrong with it: [one or two sentences]

Common mistakes and misunderstandings

Writing a vague description. "Helps with writing" will either never be chosen or be chosen for everything. The description is the only part the AI sees until it decides to open the skill, so it has to say plainly what the skill does and when to use it. Anthropic's guidance gives "Helps with documents" as an example of what not to write.

Treating a skill like a system prompt. Rules you need in every single conversation, such as "always reply in British English", belong in the standing instructions. A skill is for a particular kind of task. Putting everyday rules in a skill means they apply only when that skill happens to load.

Writing an essay instead of a recipe. Long background explanations waste the space progressive disclosure is designed to save, and bury the rules that matter. Anthropic's guidance suggests keeping the main SKILL.md under 500 lines and assuming the AI is already smart; most personal skills need a single page. If you have a long style guide, put it in a separate reference file and point to it.

Leaving out the example. Rules describe what you want; an example shows it. One good finished piece often fixes problems that five extra rules couldn't. This is the same reason few-shot prompting works.

Thinking a skill gives the AI new powers. A skill is know-how, not access. A skill for "post my weekly update" is useless if the agent has no connection to the place you post. Access comes from connectors and tools; the skill tells the agent how to use them your way.

Installing third-party skills without reading them. This is the serious one. A skill can contain instructions the agent will follow and scripts it will run on your machine or in your account. The ClawHub case shows what a malicious skill looks like: a helpful-sounding tool with a "setup step" that installs malware. Before installing anyone else's skill, open every file and read it. Be wary of any skill that asks you or the agent to download something, paste a command into a terminal, enter passwords or keys, or send information to an unfamiliar web address. Stick to sources you trust, such as the tool maker's own collection or people you know. If you run a team, Anthropic's help centre says, at the time of writing, that Enterprise administrators can switch on scanning of uploaded skills for malicious content.

Assuming a skill will always be used. The agent decides whether to pull a skill, and it can get that wrong. If a task matters, you can name the skill in your request ("use my book-blurb skill"). Check the output either way; a skill makes results more consistent, not guaranteed.

Never updating it. Your prices change, your style drifts, your newsletter platform changes its layout. A skill written in January and never touched will be confidently wrong by summer. Put a reminder in your calendar to reread your skills every few months.

Assuming skills work the same everywhere. The shared format means the SKILL.md file itself travels well, but whether a tool can run scripts, where you install skills, and which plans include them all vary. Test a skill in each tool you plan to use it with.

Related elements