Claude Code Skills: The Complete Official List

Anthropic publishes nineteen Agent Skills. That is the whole official set, and it is smaller than most lists you will find, because most lists pad the number with community repositories and count them as if they shipped with the product. This is the official nineteen, grouped, with the word count of each instruction file and a link to the raw file so you can read one before you install it.

The community collections are real and some of them are good. They are further down, kept separate, because the install route is different and so is who supports them when something breaks.

The 19 official skills

Skill	Group	Plugin	What it does	Size
xlsx	Documents	document-skills	Reads, writes and reformats spreadsheets, including formulas and charts	1,312 words
docx	Documents	document-skills	Creates and edits Word documents, including tracked changes and comments	980 words
pptx	Documents	document-skills	Builds and edits PowerPoint decks from content you supply	3,129 words
pdf	Documents	document-skills	Extracts text and tables, merges, splits, fills forms and runs OCR	1,010 words
canvas-design	Design and visual	example-skills	Produces original posters and visual art as PNG or PDF	1,749 words
algorithmic-art	Design and visual	example-skills	Generates algorithmic art in p5.js with seeded randomness	2,715 words
frontend-design	Design and visual	example-skills	Guidance for designing a UI that does not look generic	1,516 words
theme-factory	Design and visual	example-skills	Applies a consistent theme across slides, docs and reports	486 words
brand-guidelines	Design and visual	example-skills	Applies a defined brand palette and typography to any artifact	329 words
mcp-builder	Building things	example-skills	Walks through building a production-quality MCP server	1,135 words
skill-creator	Building things	example-skills	Creates, improves and measures skills, including this format	5,101 words
web-artifacts-builder	Building things	example-skills	Builds multi-component HTML artifacts with modern tooling	444 words
webapp-testing	Building things	example-skills	Drives and tests a local web app through Playwright	501 words
doc-coauthoring	Writing and comms	example-skills	A structured workflow for writing documentation with you	2,466 words
internal-comms	Writing and comms	example-skills	Writes internal announcements, updates and memos to a house format	211 words
slack-gif-creator	Writing and comms	example-skills	Makes animated GIFs sized correctly for Slack	1,103 words
claude-api	Reference and behaviour	claude-api	Reference for the Claude API and SDK: model ids, pricing, tool use, caching	14,071 words
academy-guide	Reference and behaviour	academy-guide	Points you at the relevant Claude Academy course for what you asked	1,209 words
discernment-nudge	Reference and behaviour	discernment-nudge	Adds follow-up questions after an answer so you check it before acting	1,751 words

The size column is the length of the instruction file itself. It is a rough proxy for how much the skill actually knows: claude-api at 14,071 words is a reference document, internal-comms at 211 is a short set of formats. Neither number is a quality score, but a 200 word skill and a 14,000 word skill are not the same kind of thing.

What an Agent Skill actually is

A skill is a folder with a SKILL.md inside it. The file opens with a few lines of YAML and then becomes ordinary Markdown. Claude reads the description to decide when the skill applies, then follows the body.

---
name: pdf
description: Use this skill whenever the user wants to do anything with PDF
  files. This includes reading or extracting text/tables from PDFs, combining
  or merging multiple PDFs into one, splitting PDFs apart...
---

# PDF processing
...

That is the entire format. Anthropic publishes the specification at spec/agent-skills-spec.md in the same repository.

Two consequences are worth drawing out. There is no code to run, so you can read exactly what a skill will do before you give it to an agent, which is not true of most tooling. And because the description is what triggers it, a badly written description is the usual reason a skill you installed never fires.

How to install them

The official skills install as plugins from a marketplace, not by copying files around. In Claude Code, add the marketplace once:

/plugin marketplace add anthropics/skills

Then install the bundle you want:

/plugin install document-skills@anthropic-agent-skills
/plugin install example-skills@anthropic-agent-skills

Those two cover sixteen of the nineteen. The remaining three ship as their own single-skill plugins:

/plugin install claude-api@anthropic-agent-skills
/plugin install academy-guide@anthropic-agent-skills
/plugin install discernment-nudge@anthropic-agent-skills

After installing, you use a skill by mentioning the job rather than the name. “Use the PDF skill to extract the form fields from this file” works, and so does asking for the outcome and letting the description do the matching.

Outside Claude Code

The same skills are not exclusive to the terminal. Anthropic states that these example skills are already available on paid Claude.ai plans, and that you can use the prebuilt skills or upload your own through the Claude API. So a skill you read here is portable: the file format is the same in all three places.

Which to install first

Installing all nineteen on day one is the fastest way to learn nothing about any of them. A working order that front-loads the ones that pay off immediately:

1. document-skills, if you touch spreadsheets, decks, Word files or PDFs at all. This is the bundle that most obviously replaces manual work, and pdf and xlsx are the two most people use first.
2. skill-creator, early rather than late. Once you have used two or three skills you will want one that does your job, and this is the skill that writes skills.
3. claude-api, if you build anything on the API. At 14,071 words it is the largest in the set and it exists so the model stops guessing at model ids, pricing and parameters from memory.
4. The design group, when you have a real artifact to make. These are the hardest to evaluate from a description and the easiest to judge from one output.

A caution that applies to all of them: a skill changes how the model behaves on every request where its description matches. Install a few, notice what changed, then add more. Nineteen at once and you cannot tell which one did what.

When a skill is the wrong shape for the problem

Before installing anything it is worth knowing which problems this mechanism solves and which have a simpler answer, because several things get reached for as skills that should not be.

• A one-off instruction is just an instruction. If you need something done differently this once, say so. Encoding it makes it permanent, and permanent is the opposite of what you wanted
• A preference that changes belongs in your own notes, not in an installed component you will forget you installed
• Project-specific rules belong with the project. Anything true of one repository and not another should live in that repository, where it travels with the code and applies to everyone working on it
• A genuinely repeated, opinionated procedure is the right fit. That is the case this mechanism exists for, and it is narrower than the number of available options suggests

The test that resolves most of it: would you be annoyed if this were done a different way? If yes, encode it. If you would shrug, you are adding something to remember in exchange for nothing.

One further distinction worth holding, because it is the one people most often get backwards. Something that needs to happen automatically, every time, without anyone deciding to invoke it, is not this mechanism at all: that is a hook, and reaching for the wrong one produces a component that works perfectly whenever somebody remembers it exists. If the value depends on it firing unprompted, you want the thing that fires unprompted.

When a skill earns its place, and when it does not

Installing everything available is the obvious move and the wrong one, because each addition is something the assistant has to consider and something you have to remember.

A skill is worth having when all three of these are true:

• You do the task repeatedly. Once a quarter is not a workflow, it is a search away
• The task has a right way to be done that you would otherwise re-explain each time. Encoding that is most of the value
• You would notice if it were done the other way. If any reasonable approach is fine, you did not need the skill

And it is not worth having when it duplicates something already covered, when it encodes a preference you are still changing your mind about, or when you installed it because it looked useful rather than because you had the problem.

That last case is the common one and the hardest to resist, because browsing a list of capabilities naturally generates the feeling that you might need several of them. The check that works: before installing anything, name the specific occasion in the last month when you would have used it. If you cannot name one, you are solving a hypothetical, and hypothetical problems have a way of staying installed long after you have forgotten what they were for.

One habit worth adopting early: review what you have installed every couple of months and remove whatever you have not used. Additions are easy and accumulate; removals almost never happen on their own, and an environment nobody can explain is harder to debug than one that is missing something.

There is a second reason to prune that has nothing to do with tidiness. Everything installed is context the assistant carries into unrelated work, and context is finite. A component that applies to one task in twenty is still present for the other nineteen, which is a small cost paid constantly rather than a large one paid once. Keeping the set small is not housekeeping, it is the thing that keeps the assistant sharp on the work you are actually doing today.

For how the mechanism itself works rather than what exists, the official documentation is the reference, since the details move faster than any list of them.

The community collections

Three repositories account for most of what people mean when they say there are hundreds of Claude Code skills. These are not published by Anthropic and are not supported by Anthropic, which does not make them bad, only different in who answers when something goes wrong.

Repository	What is in it	Install
anthropics/skills	The official nineteen above, plus the spec and a template	Plugin marketplace
coreyhaines31/marketingskills	Marketing skills: CRO, copywriting, SEO, analytics, ads, pricing, attribution	Clone into your skills directory
apify/awesome-skills	Scraping and data-collection skills across ecommerce and social platforms	Clone into your skills directory

For a repository that is not a marketplace, the install is a clone into your skills directory:

cd ~/.claude/skills
git clone https://github.com/coreyhaines31/marketingskills.git

Or take a single file, which is the version to use if you care what you are running:

mkdir -p ~/.claude/skills/pricing
curl -o ~/.claude/skills/pricing/SKILL.md \
  https://raw.githubusercontent.com/coreyhaines31/marketingskills/main/skills/pricing/SKILL.md

If you want the set for one job

The official nineteen are general purpose, and the community collections are large enough that finding the handful relevant to your work is its own task. We have already done that filtering for four roles, each with the skills that map to jobs those people actually do weekly and a download link for every one:

• Amazon seller skills, for pricing, listings, ads and competitor tracking
• Ecommerce skills, for storefront, conversion and merchandising work
• Affiliate marketing skills, for partner programmes and attribution
• Influencer marketing skills, for creator sourcing, outreach and campaign measurement

If what you want is the agent rather than the instructions, our guide to influencer marketing AI agents covers what is worth automating and what is not.

What these are not

Three things get promised on your behalf when people write about skills, and none of them are true.

A skill is not an integration. It carries instructions, not credentials or network access. If a job needs to reach your ad account or your database, that is an MCP server, and mcp-builder exists because the two are different layers. The packaging that can carry both, plus slash commands, hooks and subagents, is a plugin: our guide to Claude Code plugins covers the official directory of 315 and how it differs from this list.

A skill is not a guarantee. It raises the floor on a task by giving the model a method instead of improvising one. It does not make the output correct, and on anything consequential you still check the work.

More skills is not better. Each one competes for the model’s attention, and a vague description fires when you did not want it. The useful collection is small and matched to what you actually do.

Frequently asked questions

The official set is small enough to read in an afternoon, and that is the point of the format. Pick the three that match what you did last week, read the files, and install those. The rest will still be there.