What Are AI Skills? A Guide to AI Skills for Marketing Teams

Someone on a marketing team gets asked to "build us a Claude skill for briefs", and the request lands with no explanation of what a skill is. Most search results describe the format from the outside.

This guide opens the format up.

I'll start with what sits inside a skill folder and how AI tools load one. Then come public skills from Anthropic and from marketer Corey Haines, each read for the rule it's built around.

The guide ends on how to write a first skill of your own, using the shape those public examples share.

AI Skills Explained

An AI skill is a saved procedure an AI tool follows for a specific job. You write it once, install it, and from then on the tool uses it whenever a request matches.

The format comes from Anthropic, which built skills for Claude and then published the format as an open standard at the Agent Skills site. That site now documents the format for every tool that adopts it, and calls the folders Agent Skills.

Inside Claude, they're usually called Claude skills.

What a skill gives a marketing team is consistency without repetition:

That last row is the one that changes how a team works. A prompt improves only for the person who keeps editing it, while a shared skill improves for everyone who uses it the next morning.

What's Inside an AI Skill Folder

A skill is a folder with one required file, SKILL.md. The file opens with a short block of metadata, a name and a description, followed by instructions in plain text.

Everything else in the folder is optional. Anthropic's guide to creating custom skills uses a brand-guidelines skill for an invented company, Acme, to show one of each kind of file:

brand-guidelines/
├── SKILL.md          # required: frontmatter and instructions
├── references/       # optional: documentation Claude reads when a step calls for it
│   └── voice-and-tone.md
├── assets/           # optional: templates and files Claude copies or fills in
│   └── slide-template.md
└── scripts/          # optional: code Claude runs while following the skill
    └── check_contrast.py

Each folder has a separate job. The reference file holds Acme's writing rules, the asset is a slide outline to fill in, and the script checks whether a text-and-background color pair has enough contrast to read.

Code only belongs in a skill when something has to pass or fail a check, and Anthropic's advice is to start with plain instructions and add scripts only when a step needs one.

The instructions in SKILL.md can include step-by-step procedures, input and output examples, templates and edge cases to handle. Anthropic suggests keeping the file under 500 lines and moving longer material into the reference folder.

How AI Skills Work When You Ask for Something

A skill doesn't sit in the conversation the whole time. The Agent Skills format describes three stages, which it calls progressive disclosure:

1. Discovery. When the tool starts, it reads only each installed skill's name and description.
2. Activation. When your request matches a description, the tool reads that skill's full SKILL.md.
3. Execution. The tool follows the instructions and opens the reference files or runs the scripts when the instructions ask for them.

Say someone asks Claude to "make the Q4 planning deck and apply our brand guidelines" with a brand-guidelines skill installed. So far, Claude has read only the description. That description mentions presentations and brand guidelines, so Claude opens the full file and follows it.

The same order explains how a team can install dozens of skills without slowing anything down. Until a task calls for one, each skill costs the tool a couple of lines of reading.

It also puts a heavy load on the description. If the description doesn't match what someone types, the tool never opens the rest of the file, no matter how good the instructions underneath are.

Claude Skills and the Other Tools That Read Them

Claude skills and Agent Skills are the same format. Anthropic built it for Claude. Because the standard is open, you can run a skill written once in any other tool that supports it.

The client list on the Agent Skills site includes these, among many others:

Most of that list is coding tools, which shows where the format took hold first. For a marketing team, Claude is usually the place to start. Its apps let you turn a skill on under Customize, then Skills, without opening a terminal.

The coding tools still matter to marketers who work with developers. A brand skill written for Claude can run inside the same code editor a developer uses to build landing pages, so the rules travel with the work.

Portability has a limit to plan around. The folder travels between tools. A skill that tells the tool to call a specific connector only works where that connector exists.

The GTM Exchange already has a guide to Claude skills for marketers, covering how to install them. This piece stays on what a skill is and what a good one looks like.

AI Skills for Marketing: Public Examples to Learn From

Anthropic keeps a public skills repository of working examples, and its guide points to four of them as models for the common shapes. Marketer Corey Haines publishes a marketing collection called marketingskills under an MIT license.

I read the ones most useful to a marketing team to see what each does that a prompt can't. Each turned out to be built around a single idea a draft can be checked against.

Brand Guidelines, the Instructions-Only Shape

Anthropic's brand-guidelines skill is a single SKILL.md with no scripts or reference files. The guide calls it a good model for a skill that only provides instructions, such as a style or formatting rule.

The Acme version in the guide shows the content: brand colors as hex codes, fonts for headers and body text, sizes for each heading level, logo rules for light and dark backgrounds, and a list of the documents the rules apply to.

That last list is the part a prompt usually leaves out. It tells Claude when the rules apply, so a quick internal note doesn't come back dressed as a client deck.

What Still Needs a Person

Your own brand. Anthropic's published skill applies Anthropic's styling, and its guide advises copying a skill's structure and replacing the contents with your own.

Internal Comms, Instructions Plus Examples

Anthropic's internal-comms skill drafts status reports, newsletters and FAQs. Its folder adds an examples folder of sample documents, which SKILL.md tells Claude to consult.

Examples carry what's hard to describe in a rule. A sample status report shows length, tone and section order in one go. Claude can match it more closely than it could follow a paragraph describing the same thing.

What Still Needs a Person

The facts in the update. The skill shapes the document, while what happened this week still has to come from someone who knows, and the examples need updating when the team's formats change.

Cold Email, Built on the Delete-the-Opening Test

Corey Haines's cold-email skill writes B2B cold emails and follow-up sequences. Before writing, it reads a product-marketing file if your project has one, so it doesn't ask for positioning it already has.

Its central rule is a test you can run on any draft.

Remove the personalized opening line and read what's left. If the email still makes sense without it, the personalization wasn't connected to the problem, so the email needs rewriting.

✗  "I hope this email finds you well. I came across your profile and..."
✓  An opening observation the rest of the email depends on, so removing
   it breaks the argument.

The first line comes from the skill's own list of phrases to avoid. The rest of the file follows the same habit, with a quality check before showing any draft and a list of things never to send, such as fake "Re:" subject lines.

Its references folder carries five files: benchmarks, a personalization system, subject-line data, follow-up sequences and copywriting frameworks. SKILL.md stays short and points to each one at the step that needs it.

What Still Needs a Person

The signal and the proof. The skill asks for a reason to write to this person and a result you can back up. It writes from whatever you give it, so a weak signal still produces a weak email.

Skill Creator, a Skill for Testing Skills

Anthropic's skill-creator helps you write a skill and then check whether it improves Claude's output. It runs your skill on test prompts you agree on and shows the results.

In Claude Code, it goes further: it runs the same prompts without your skill as a baseline, then reports pass rates, timing and token counts so you can compare the two.

That baseline answers the question a few test prompts can't. A skill can load every time and still produce work no better than a plain request.

What Still Needs a Person

Deciding what a pass looks like. The tool runs the comparison, while you provide the prompts and the standard.

Where to Find Public AI Skills

Several public libraries are a good first stop before writing anything from scratch:

Anthropic marks the skills in its repository as demonstrations, and advises testing them in your own environment before relying on them for critical work. That advice applies to every public library.

A skill that hasn't been updated in a long while may describe a tool's old settings, so check when it last changed before you install it.

What Good Public Skills Have in Common

Read side by side; these skills share a shape I'd copy for any new one.

First, each names the job in a sentence a stranger could follow. Second, each states its standard as something a draft passes or fails. "Make it personal" became the delete-the-opening test, and "keep it on brand" became a list of hex codes and fonts.

Each also says when it applies or what it leaves to other skills. The cold-email skill sends lifecycle email to a separate emails skill and list building to a prospecting skill, so two skills don't fight over one request.

The order inside each file follows the order a person would need it: the job first, then the rules it can't break, then the steps.

Reference material sits outside the main file until a step asks for it, which keeps SKILL.md short enough to read in one sitting.

Anthropic's own best practices point the same way: keep each skill focused on one workflow, write a specific description, start with instructions, show the output you expect, and test after each change.

The Line That Decides Whether a Skill Runs

I expected the instructions to be the hard part of writing a skill. Reading these files, the description at the top does more of the work. It's the only part Claude reads before deciding whether to load the skill.

Here's the frontmatter from Anthropic's Acme example:

---
name: brand-guidelines
description: Apply Acme Corp brand guidelines to presentations and documents, including official colors, fonts, and logo usage.
---

It names the job and the kinds of documents it covers. A description that only said "brand stuff" would miss most of the ways someone asks for a branded deck.

The cheapest test for a description is to ask three colleagues how they'd request the job, in their own words. Any phrase the description doesn't cover is a request the skill will miss.

I'd also add a line saying when not to use the skill and which one to use instead. Without it, two skills with similar descriptions can both claim a request, and you can't tell which one the tool picked.

The name has rules too. It accepts only lowercase letters, numbers, and hyphens, up to 64 characters, and it must match the folder's name. The description can run up to 1,024 characters.

When a Prompt Is Enough and When to Write a Skill

Plenty of jobs are faster as a prompt. A one-off request, such as rewriting a single email or summarizing a call, takes less time typed straight into the chat.

Anthropic's guide names two more cases where a skill is the wrong tool. Live data from another app needs a connector, and instructions you want in every conversation belong in your personal preferences or project instructions.

A skill starts paying for the effort when three things are true at once. The job comes up every week, more than one person does it, and someone can check the output against a right answer.

Briefs, brand checks and recurring reports usually pass all three tests. A launch announcement for a product you'll only launch once usually doesn't, however long the prompt behind it runs.

How a Skill Differs From a Prompt, a Project and a Connector

Skills sit alongside other ways of giving an AI tool context. The lines between them blur in conversation:

MCP stands for Model Context Protocol, the open standard for connecting AI tools to other apps: a connector gives the tool access, and a skill tells it what to do with that access.

Anthropic's guide gives an example of the two working together. A skill can teach Claude how your team uses an issue tracker you've connected, such as which project new issues go in and which labels to use, while the connector does the reading and writing.

The lead enrichment workflow on the GTM Exchange is a marketing version of the same split, with a skill holding the rules and connectors reaching Apollo and HubSpot.

What It Takes to Write a Good AI Skill

Reading these files side by side shows what each part of a skill asks of the person writing it:

None of this is new to marketing, since it's what anyone does when they write a good brief or a style guide, applied to a reader that follows instructions literally.

A clear picture of a bad result is the one I'd start from. Each example above depends on someone being able to say what a bad version looks like.

A cold email whose opening could go to anyone and a deck in the wrong colors are both failures a person has to recognize before a file can rule them out.

The reason behind each rule comes next. It lets the tool, and the next person who edits the file, tell whether a new case falls under the rule.

The words people use are the part I underrated the longest. The person who wrote the process knows its proper name, while everyone else asks for it in their own words, and the description has to catch both.

A trusted source for each fact sounds like the dullest of the five. It decides whether a skill stays accurate. A brand skill pointed at the current brand file keeps working when the palette changes, while one with old hex codes goes stale.

Our Claude SEO content workflow shows the same thinking at a larger scale, where a chain of steps checks facts that a single prompt let through.

Writing a Skill Is Different From Writing a Prompt

Good prompts and good skills look alike on the page, but they require different habits. You write a prompt for one conversation, so it can lean on what you'll say next.

Your skill has to work for someone you'll never meet, on a request you didn't see coming. That pushes you to write down the decisions you'd normally make in your head, such as what counts as done and which sources to trust.

Marketers who write briefs for freelancers already have most of this habit. The difference is that a freelancer asks questions when a brief is vague. A tool usually guesses, and then fills the gap with whatever looks most likely.

What to Check Before Installing Someone Else's Skill

Skill libraries are easy to find now, including the Claude skills directory on the GTM Exchange. Before adding one from a stranger, I'd read four parts.

Start with the description, which tells you which requests will trigger the skill. One written too broadly will grab requests you meant for something else, and you'll only notice when the output looks wrong.

Read the hard rules next. They show what the author was worried about. That's usually a good guide to where the skill is strong and where it was never tested.

Then look at what it fetches and where it writes. A skill that reads your CRM or posts to Slack through a connector can change data, so I'd want to know exactly which actions it takes.

Last, open any scripts folder. A skill can include code the tool runs on your machine. I wouldn't install one whose scripts I couldn't read myself or have someone check.

If a skill passes those four reads, I'd still run it on a test request before trusting it with live work. Its first output tells you more than its description, especially about tone.

Where AI Skills Still Fall Short

Skills are useful, and they come with limits I'd want any team to know before building a library:

• A skill is only as good as the process behind it. Writing down a bad process gives you the same bad result every time.
• Descriptions can miss. A skill whose description doesn't match the request never loads, and nothing tells you it was skipped.
• Two skills can claim one job. Without a line pointing elsewhere, the tool picks one, and you won't know which.
• Instructions don't check themselves. A rule depends on the tool following it, and only a script or a person can confirm it did.
• Reference files go out of date. A brand file or a pricing list written months ago carries the old facts forward without warning.
• Support varies by tool. The format is shared, while install steps and the extras a skill can use differ from tool to tool.

Writing Your First AI Skill for Marketing

If I were starting a library from nothing, I'd pick one job our team does the same way every week. A brief, a recurring report and a brand check all fit. Each already has a right answer to compare against.

Anthropic's guide suggests starting with one sentence describing the task and what a good result looks like. That sentence becomes the description, and the rest becomes the instructions.

The starter skill below follows the shape the public examples share. Each part has a job:

1. The name. Lowercase words joined by hyphens that describe the action, like applying-house-style, matching the folder's name.
2. The description. What the skill does, the phrases people type when they want it, and when to use something else.
3. The hard rules. Each one with the mistake it prevents, so a colleague can tell whether a new case falls under it.
4. The steps. In order, as a checklist the tool works through.
5. What still needs a person. The part of the job the skill hands back, stated plainly.

Fill in the brackets with your own house rules. Zip the folder itself, so the archive holds the folder and SKILL.md sits inside it, then upload it and turn it on under Customize, then Skills.

Test it on three requests worded the way a colleague would type them.

If any of the three doesn't load the skill, the description needs more of the words people use. If it loads and the output breaks a rule, the rule needs a clearer test, one a draft can visibly pass or fail.

Expect a few rounds of that before the skill behaves the same way every time. Once it does, skill-creator can show whether it improves on a plain request.

These pages on the GTM Exchange help with the next one:

• The Claude skills directory collects marketing skills you can install, each with its limits named.
• The product-marketing skill writes a positioning file that other skills read first.
• The best Claude skills for marketing are tested and sorted by job.
• The prospecting skill builds the list the cold-email skill writes to.
• The emails skill designs a sequence around your product, from welcome to win-back.
• The churn-prevention skill covers cancel flows and failed-payment recovery.
• The content brief skill is a worked example of a skill run end to end.
• The fact-checking skill workflow checks each claim in a draft against a fetched source.
• The LinkedIn outreach skill pairs instructions with a script that enforces the limits.
• Shiyam Sunder's Claude skills write-up walks through a set of marketing skills node by node.