Skills for Claude Code: what they are, how to install them and where to find ready ones
A skill is a folder with a single file describing how to perform a specific task properly. The agent loads it itself when the task matches.
A skill is a folder with a single file describing how to perform a specific task properly. The agent loads it itself when the task matches.
Put simply: you describe the procedure once, and from then on the agent follows it instead of improvising afresh every time.
Below: how a skill differs from a prompt, how to install a ready one in three ways, how to write your own in five minutes, what to check in someone else's skill, what people are paid for this, and where to find vetted ones — the catalogue holds over 10 000 with open descriptions.
The guide is live: the skill format is still changing, and this piece follows the releases.
The short version
A skill is not a prompt but a file containing a procedure. A prompt lives once; a skill is picked up automatically when the task matches.
It installs in a minute: a folder with a SKILL.md file goes in the right place and the agent sees it.
It is written in five minutes. There is essentially one mandatory field — the description the agent uses to decide when to switch the skill on.
It does not consume context. Only a short signboard stays in the agent's memory; the procedure itself loads at the moment it fires.
Read someone else's skill before installing as you would read someone else's script: it can run commands.
If you do not know which skill you need — the Cerebrum section breaks the task into five to seven steps and assembles the chain itself, from a base of over ten thousand skills.
There is money here. A finished commercial proposal for such a chain takes 5–7 minutes to assemble; the $1 900 proposal from a real deal was put together exactly this way.
Part 1. What a skill is
A skill is a way to package a repeating instruction into a file that the agent loads only when it is needed.
According to the official Claude Code documentation, creating a SKILL.md file with instructions is enough — Claude adds it to its set.
When it is worth creating one. If you keep pasting the same multi-step procedure into the chat. Or if the section describing your project has grown from a couple of facts into a full instruction.
How it is built. A short settings block at the top, ordinary instruction text below. Of the settings, essentially one line is mandatory: the description, which the agent uses to decide when to switch the skill on. The command name comes from the folder name.
Skills follow the open Agent Skills standard, which Anthropic introduced in late 2025. Within months it was picked up by Cursor, Codex, Gemini CLI and dozens of other environments.
How a skill differs from a prompt, MCP and a subagent
Four things people constantly confuse. The difference is in purpose:
A prompt is one-off request text. In the next session it is gone. A skill lives as a file and is picked up on its own.
A project description file (CLAUDE.md) holds permanent facts always present in the agent's memory. A skill is a procedure loaded on demand. That is why bulky checklists get moved out of the project description into a skill, so as not to burn memory for nothing.
MCP is a standard for connecting external tools: databases, services, storage. A skill is the knowledge of what to do with those tools. They do not compete: a skill can describe how to use MCP tools properly.
A subagent is a separate instance of the agent with its own memory for a subtask. A skill can be run inside such an instance, but a skill in itself is an instruction, not a second agent.
In practice they add up to one loop. Permanent facts about the project sit in the description, access to services comes from MCP, the knowledge of "how to do this properly" is packaged into a skill, and where necessary the skill moves the heavy autonomous part out into a separate instance.
So the question "skill or MCP" is put wrongly: these are different layers and they work together.
How the agent knows when to switch a skill on
By the description. In an ordinary session the agent sees only the short signboards of all available skills — that is cheap, on the order of a hundred words each. The full procedure is pulled in only at the moment the skill fires.
Hence the main rule: the description must state explicitly what the skill does and when to apply it. Not "working with text" but "use when the user asks to review a diff or write a commit message". This is the one field Anthropic recommends always filling in.
The second route is calling it manually: type / and the skill name. Useful for procedures you want to keep under control: long checks, deployment, report generation. Some skills the agent switches on itself, others run only on an explicit call.
Part 2. How to install one
Every method comes down to the same thing: a SKILL.md file ends up in a folder the agent reads. The only difference is whether you put it there by hand or the agent does it for you.
The simplest way: ask the agent
If folders, paths and the command line put you off, you can leave them alone entirely.
Open Claude Code and write in ordinary words: here is a link to a skill, install it and switch it on. The agent downloads the file, puts it where it belongs and confirms the skill is active. From then on you simply set the task and it picks up the procedure itself.
You can also just attach a downloaded SKILL.md file to the chat and ask for it to be installed.
This is not a simplified or cut-down version: the result is exactly the same as installing by hand. The path is typed by the agent rather than by you. What follows is how it works underneath, if you want to understand it or hand one set to a whole team.
By hand
Create the skill folder in one of the places Claude Code looks at:
Personal — ~/.claude/skills/<name>/SKILL.md, available across all your projects.
Project — .claude/skills/<name>/SKILL.md, only in that repository. Convenient to commit so the whole team gets one set.
From a plugin — wherever the plugin is enabled.
Where names coincide, the enterprise level overrides the personal one and the personal overrides the project one. Claude Code watches the folders and picks up changes within the current session, with no restart.
A file downloaded from the catalogue simply needs unpacking into .claude/skills/<name>/ — the agent will see it on the next matching task.
A subtlety with large repositories. Project skills load from the starting folder and every parent folder up to the root, while nested ones connect at the moment the agent first touches a file inside such a subfolder. This lets an individual package carry its own set, switched on exactly when you are working with that package.
From a marketplace
Install a skill as part of a plugin. The official marketplace is added with /plugin marketplace add anthropics/claude-plugins-official, after which a plugin installs like this: /plugin install skill-creator@claude-plugins-official.
Plugins are convenient because alongside skills they bring agents, hooks and service connections in one package.
By a command from the catalogue
Many catalogues give a ready installation line of the form npx skills add <owner>/<repo> --skill <name> --agent claude-code. It places the required file in the project folder.
How to check it is in place
Type / and the skill's folder name. If it appears in the suggestions, the agent has read it.
The second test: ask a question matching the description. The agent should offer or perform the procedure by itself.
If it does not appear, check that the folder is in the right place, contains a file named exactly SKILL.md, and that the settings include a clear description.
Skills across a team
Project skills are the way to give everyone the same set of procedures: put the folder in the repository, commit it, and everyone who clones the project has the same commands.
Convenient for shared review rules, deployment and code generation: instructions do not need sending round a chat — they travel with the code. Updates reach the team with an ordinary git pull.
Part 3. How to write your own
In five minutes. Create a folder and put a SKILL.md in it with a description and instructions.
A minimal working example from the documentation:
```
description: Summarises uncommitted changes and highlights risks. Use when the user asks what changed, wants a commit message, or asks to review their diff.
Current changes
!git diff HEAD
Instructions inside SKILL.md
Summarise the changes above in two or three points, then list the risks: missing error handling, hardcoded values, tests that need updating. If the diff is empty, say so.
```
The line ` !git diff HEAD ` is live data substitution: Claude Code runs the command and replaces the line with its output before the agent reads the skill. The instructions arrive with the current content already in place.
What it consists of
A skill is a folder with SKILL.md as the entry point; templates, examples and scripts can sit alongside it. The permitted settings are name, description, license, compatibility, metadata and allowed-tools — all optional, with only the description mandatory in practice.
The description is both signboard and trigger. Write what the skill does and when to apply it, in the words a person would use to phrase the request.
The body is the procedure itself: a step-by-step checklist, not abstract wishes. A short verifiable body works more reliably than a long argument.
Scripts and templates — move repeatable mechanics into separate files referenced from the body, so the bulk stays down. They load only when genuinely needed.
Why structure works better here than phrasing
A skill is structured text, and that is not a coincidence.
Claude is trained to read markup as structural boundaries. Put simply: if you mark parts of a request as separate blocks, it understands where the instruction ends and the material begins, which part is an example and which sets out the requirements for the result.
That is exactly why a skill with a clean separation into description, body and referenced files works more reliably than one written as continuous prose — even with identical content.
And the same holds beyond skills. The technique applies to any working request: break it into blocks and you get a different class of answer. It shows most on long documents, where the instruction usually dissolves into the material.
The express course "Claude: from your first prompt to your own library of mega-capabilities" covers exactly this: how this model differs from the rest, structured prompts with markup for business tasks, and twelve ready mega-prompts for working scenarios.
It is useful specifically as groundwork for skills: understanding why structure works, you write the body of a skill deliberately rather than copying someone else's template.
$9, included in the Creator plan alongside the main training. Not part of the three free Basic days.
It is written in iterations
A good skill does not come out right first time. The working cycle: write a minimal version → call it on a real task → see where the agent went wrong → adjust the description or a step.
The official marketplace provides a dedicated tool for this — the skill-creator plugin, which helps generate and calibrate descriptions.
Editing as you go is convenient: Claude Code sees changes within the current session — save the file and the next call already uses the new version.
Before writing your own, look at the advanced ones
There is a part of the catalogue people usually do not know about.
The "Vibe coding" category holds over 1 200 advanced skills (1 297 in the database) — procedures for specific development tasks, the kind that normally take weeks to assemble.
What they cover by technology:
Payment systems — Stripe: subscriptions, one-off payments, billing.
Mobile development — Android in the native and modern stack, iOS in Swift and the Apple ecosystem.
Mini apps — the WeChat Mini Programs ecosystem, front end and back end.
Desktop — Electron, Qt and C++.
Browser extensions — the Chrome Extension API.
Web3 — Solidity and smart contracts.
What gets built end to end on this:
- A SaaS platform — a web service with sign-up and login, text generation, plans, Stripe payments and an admin panel
- A testing service — question generation, a sitting interface, a management system
- A mini app with a back end, a database and business logic
- Native applications for Android and for iOS
- A PWA — a web application with offline mode, push notifications and device installation
- A browser AI assistant as an extension
Three typical mistakes
A vague description. The agent does not understand when to call the skill, and it "does not fire". The most common problem by far.
A wall-of-text body. The longer and more abstract the instruction, the more often the model prefers its own route.
Permanent facts inside a skill. Those belong in the project description. Keep skills for procedures.
If a skill seems to stop having any effect after the first answer, the content is usually intact — the model is simply choosing other tools. Strengthen the description or call the skill again after the context has been compacted.
Part 4. What to check in someone else's skill
Someone else's skill is executable instructions for your agent. Read it in full, the way you read someone else's script.
Commands in the body. Lines of the form ` !... run before the agent reads the skill. Make sure what is there is harmless, like git diff`, rather than downloading and running something from the network.
Requested permissions. A skill can ask for access to tools. Do not grant extended rights to something that does not need them for the task.
Scripts in the folder. If separate files sit alongside, read them: the body can run them without asking separately.
Provenance. Prefer skills with a clear origin. In the catalogue the description is visible before download — which lets you judge a skill before the file reaches you.
Part 5. Where to find ready ones
The Skills for Claude Code catalogue — over ten thousand skills with open descriptions in two languages. Each has a card at its own address, and the files download with a subscription.
Descriptions are open to guests — you can study what a skill does before registering at all. The catalogue is fed by daily radars, so fresh finds appear without manual curation.
Why this beats a list of links: every skill has a card with a clear description and a category, rather than a bare repository link that may move within a month.
The express course, if you would rather go in order
Skills are part of the course "Claude: from your first prompt to your own library of mega-capabilities", $9.
A dedicated section covers exactly this: what a SKILL.md folder is, how it loads only when needed, direct access to 500+ community skills by category — brand, marketing, code, e-commerce, content — and the built-in skill-creator, which writes a SKILL.md for your task if no ready one exists.
Plus the thing without which skills work at half strength: the eight tags of a mega-prompt and a breakdown of what each is responsible for. The skill sets the procedure; the structure of the request sets the quality of the result.
How to search properly
Skills are arranged by task: engineering and code review, working with documents, content generation, analytics, marketing, operations and others.
Search from your task, not from the ratings. A skill is worth exactly as much as its description matches what you actually do. A skill with an accurate description will be switched on reliably; a fashionable one that has nothing to do with your work simply will not fire.
Where to start. The easiest entry is the free in-house content generation skills — there are twenty of them, they need no key configuration, and they show the mechanic itself: a skill leading an agent through steps. Available to any signed-in user.
After that, go by your stack: code review, working with documents, analytics.
Part 6. If you do not know which skill you need
The honest problem with a catalogue of more than 10 000 positions: to choose the right one you have to roughly understand what you are looking for. And a complex task is usually solved not by one skill but by several in sequence — and which ones exactly is not something an outsider knows.
There is a separate section for this — Cerebrum.
What it does
You describe the task in ordinary words. Not "find me a skill for Stripe" but "I want a service with subscriptions and a user account area".
From there it:
Asks a question if the task is vague. One or two, to the point: what is already in use, where the data comes from, what the output should be. Not a questionnaire, but the minimum without which the selection would be guesswork.
Searches by meaning rather than by words. One query across two bases at once: over ten thousand skills and more than 800 deployment-ready open source projects. The phrasing "I want emails to sort themselves into folders" finds the right thing even when those words appear nowhere in the tool's description.
Assembles one chain rather than a list of options. The task is broken into five to seven steps in execution order: step 1 is this skill, here is its role and what it passes on; step 2 is the next one, and so on to the result. One position per step, with no "here are twenty options, work it out".
Explains the joins. For each step, how to obtain it and how to connect it to the neighbouring one. For the open source parts it produces a ready prompt for Claude Code: what to clone, what to read, what to build, what to adapt. Copy, paste, go.
Tells you which step to start with today. Not "good luck" but a specific first point of entry.
Why a chain rather than a list
A list of tools anyone can find. The problem is always the other one: what connects to what, in what order, and what passes to whom.
And skills and open source get mixed into one chain by the logic of the task rather than by catalogue section: in one place step one is a skill for Claude Code and step two is a service you host yourself. Elsewhere it is the other way round.
Three things that are non-negotiable here
Links only to positions that actually exist. Cerebrum works from what the search returned and does not invent projects or skills — that is written into its rules.
No invented deployment method. If a project's description does not say how to bring it up, it says so: see the README — rather than a plausible fabricated instruction.
Nothing unsuitable gets mentioned. The answer is not padded for length.
What it costs
Three levels, chosen in the input field according to the difficulty of the task.
| Level | For what | Turn ceiling |
|---|---|---|
| light | selection, short tasks | up to $0.60 |
| medium | complex chains | up to $1.00 |
| high | the hardest tasks | up to $2.00 |
The ceiling is not the price. It is the upper limit reserved at the start; after the answer the actual spend is calculated and the difference returns to the balance automatically. A typical turn costs noticeably less than the ceiling, and an empty answer means a full refund.
And a rule built into the behaviour: no paid tool runs without your agreement. The model and the price are named first, then it waits for confirmation.
Who this replaces the catalogue for
People who do not know what to search for. They have the task but not the tool names.
People with a complex task. Five to seven steps are not assembled by hand out of more than 10 000 positions.
People who do not want to work out the joins. The ready prompt for Claude Code removes most of the connecting work.
Who does not need it: if you know exactly which skill you want, open the catalogue directly — that is faster.
Part 7. What a skill actually gives on a live task
Taken from something done in a real deal — that way the difference between a prompt and a chain is visible.
The task
After a call with a client, a commercial proposal is needed. Not as text but as a finished deck that can be sent.
What existed beforehand
The call recording had been gone through: what the client said about his losses, what he is afraid of, which lever matters most to him. Three pains, named by him. That is the content of the future proposal — it is not invented.
How it was assembled
Step one. The sales assistant received the inputs and produced a detailed brief for the deck: style requirements, emphases, colours and fonts, number of slides, structure. And separately — which phrases to avoid. A list of prohibitions is what saves you from the clichés proposals usually fall apart on.
Step two. The same assistant then checked its own answer against its own criteria. Plus a separate mode: the model took the client's role and walked the deck slide by slide — where he would tense up, what he would skip. After which it made the corrections itself.
Step three. The finished brief went to Claude Code with the presentation skill. An editable file in 5–7 minutes.
Under ten minutes for all of it.
Why a skill rather than a prompt
The presentation skill is not one request but a chain of roles: a copywriter writes the text → a designer sets the visuals → the next lays it out across slides → the final one checks the work of the first three and produces the result.
Four steps instead of one. And the fourth is the one that never happens in manual work: nobody checks their own draft against criteria.
What matters here
Not the speed. Ten minutes is a pleasant consequence. A proposal assembled in ten minutes out of anything other than the client's own words is simply quickly made rubbish.
The order. Analyse the conversation → brief → check → assemble. Each step consumes the previous result, and the output is a document in which the client recognises himself.
This proposal was paid in full — $1 900 for an annual system. The client is Zakhar Fomin, owner of a dental clinic, and the contractor is Denis, the founder; the client agreed to be named along with the figures.
Part 8. How people earn from this
The catalogue by itself is a warehouse. What sells is not the warehouse.
An industry bundle
What people buy: a selection for a specific profession. Ten to fifteen skills "for a lawyer", "for a marketplace seller", "for a clinic", tuned to that field's processes and documents.
The client is not paying for files. He is paying for somebody having chosen, configured and explained how to use them.
Why it works. Someone opening a catalogue of ten thousand positions for the first time closes it within ten minutes. The same person handed a folder of twelve with the note "this is for your tasks" starts using them the same day.
You sell the removal of choice, not the files.
The order of prices
From a scrape of 5 986 jobs over four months: a system or a pipeline runs $490–1 460, against a median one-off job of around $60.
An honest caveat: an explicit budget appears in 17% of listings, and by individual type of work there are between twenty and three hundred observations with a figure.
The second route — your own projects, faster
A proposal in ten minutes instead of an evening. A deck, a report, a document from a template. This is not about selling a service but about getting more done in the same day.
The limits
A skill does not replace substance. The most careful chain will assemble a beautiful nothing if you do not know what your client is afraid of. Going through the conversation is a mandatory part, and it does not automate.
The result gets checked. Figures, names and links are verified against sources. A skill speeds up production; it does not remove responsibility.
Not everything is worth automating. A task done once a quarter will not repay the selection and setup.
Part 9. What to open for your own task
Skills for Claude Code. Over 10 000. Cards and descriptions are open to everyone without registration. Files come with the Basic plan. Advanced ones with the Creator+ plan, or $9 one-off, or 1 000 experience points earned by working on the platform.
Ready assistants — the thing that supplies the substance. Over 200 across different categories. A skill assembles the document, but the brief for it is written by an assistant. The cycle described above needs two: the sales assistant — for analysing the call, writing the brief and running the check through the client's eyes — and one specific to your field. Basic plan.
Freelance 2.0 — if you need the whole system. 62 prompts across 11 stages of the funnel: research, positioning, content, lead magnets, social, monetisation, sites, funnels, sales with scripts and practice partners. The key property is the same as a skill's: each result feeds the next. $9, included in Basic.
Cerebrum — if you do not know which skill you need. Describe the task in words and get a chain of five to seven steps with the joins explained and a ready prompt for Claude Code. Wallet payment on actual usage, with a turn ceiling from $0.60.
Your own skills and assistants for a company — Creator+ plan.
If Claude Code is still foreign territory
Skills are a tool for people already working with an agent in code. If you came here because you heard about Claude Code but do not write code, the order is different.
The course "VIBE CODING" (Creator plan and above; the first two lessons open from Basic) takes you from nothing to your own application. Three of its modules matter for this article.
The "Introduction to vibe coding" module — how an application is actually built. What any application is made of, what systematic vibe coding is, and a walkthrough of a typical application before you build your first one from scratch.
This is what a code generation skill needs to stop being a black box: otherwise you get a result and do not understand what you got.
The "Building the prototype of your first application" module. Applied work: an idea becomes a clear technical brief, then the design, a preview of the interface and the "skeleton" of the first application.
The module "How any application works inside. How to add features, typical mistakes and methods of effective vibe coding". The architectural brief, screen logic, mockups, the final prompt, MVP acceptance — that is the level at which skills stop being someone else's files and become your working tool.
Plus what rarely makes it into technical courses: the "App Value" lesson — what an application's value is made of and what users are ready to pay for. And selling finished work is covered in the "Sales" course (Basic plan): where to find orders, how to write to clients, the first call, presenting the work and the upsell.
The order is simple. If you do not understand what the agent built, start with the course. If you already work in Claude Code and want it to follow your procedure, skills and the Claude express course are enough.
Alongside: the automation templates section — over 3 000 ready scenarios if the task is mechanical; the AI agents hub — if you need a process rather than a document.
What to do after your first skill
The ecosystem moves fast: fresh releases and techniques are covered in the daily AI digests. If you work with open models, look into the Best open LLMs 2026 hub; if you need ready processes without code, the automation templates.
FAQ
How does a skill differ from the project description?
The project description holds permanent facts that are always in the agent's memory. A skill is a procedure loaded only at the moment of use.
The practical rule from the documentation: if a section of the description has turned from a fact into a multi-step instruction, move it out into a skill. A skill's body costs no memory until it is needed.
Do skills work in the web version and in Claude Code at the same time?
Yes: they are built on an open standard supported by different environments. The file format is one. Claude Code adds its own extensions — call control, running in a separate instance, live data substitution — while the basic settings are the same everywhere.
What does using skills cost?
The mechanism itself is free — these are files your agent reads.
In the catalogue every skill's description is open at no cost, downloading the files opens with a subscription, and twenty in-house content generation skills are free to any signed-in user.
On model spend a skill is economical: only a short description stays in memory permanently, and the procedure loads only when it fires.
How many skills can be connected at once?
There is no hard limit: only short descriptions are held in memory, so a large set costs little and only the one that fires gets expanded.
In practice the constraint is not the number of skills but the quality of the descriptions: the more precise the description, the more reliably the agent picks the right one out of many.
How do I update an installed skill?
Replace the files in its folder. Claude Code watches the folders and picks up changes within the current session, with no restart.
If you have created an entirely new skills folder that did not exist when the session started, then a restart is worth doing.
Can a skill be run in isolation?
Yes: add context: fork to the settings and the skill runs in a separate instance with its own memory — the body becomes the instruction guiding it, and your conversation history is not passed along.
Useful for long autonomous tasks such as research, which should not clutter the main conversation. By default such a skill runs in the background and the result comes back on completion.
Can you earn from skills?
Yes, but what sells is not the file — it is the selection and the setup. An industry bundle for a specific profession, ten to fifteen skills tuned to its processes, is a complete service.
From our bot's job feed — 5 986 over four months — a system or a pipeline runs $490–1 460 against a median one-off job of around $60. The client pays for choice being removed: on his own he will not pick from ten thousand positions.
Try it
Registration is free. Three days of the Basic plan switch on with one button, no card needed: Basic includes 200+ AI assistants, prompt texts, skill files, five programme courses (including “Sales”) and six Basic mini-courses — among them “Freelancer positioning”, Freelance 2.0, Perplexity and Grok.
One action right now. Take a document you produce regularly and always slowly: a proposal, a report, a deck. Find a skill for that type in the catalogue and run it once on a real task.
Look not at the speed but at what your manual version usually lacked. Normally that is the final check against criteria — the step a person never performs on their own work.
What to take on the platform for this task
- Creating and updating Agent Skills — a skill that helps you write and update your own skills; advanced (Creator+ / $9 / 1000 XP)
- Interactive SKILL.md creation for Claude Code — a dialogue that turns a raw idea into a finished SKILL.md; advanced (Creator+ / $9 / 1000 XP)
- Skill Forge — a skill builder for Claude Code — building skills to the open Agent Skills standard; advanced (Creator+ / $9 / 1000 XP)
- Claude Code — an AI coding agent for the terminal — the agent itself, where skills are installed; the description is open to all, the repository link from Creator
- Programming mentor — an assistant that teaches programming from zero; from Basic
- Inside Sales Representative — sales consulting for B2B and B2C; from Basic
- The "Sales" course — the "First order" series: from finding a client to presenting the work; Basic plan
What to read next
How Claude differs from other models — why structure works better than phrasing.
What prompt engineering is — sixteen methods in four groups.
Six fields of a brief — what to supply so the output is not a template.
One assistant against three — when a chain of roles works.
A commercial proposal in ten minutes — the full breakdown of assembling a deck.
What vibe coding is — if you came for Claude Code but do not write code.
Application architecture as a building — so you understand what the agent built.
Why a project falls apart on the third edit — the description as the source of truth.
What of your builds actually sells — eight formats with prices.
AI for developers — three ready assistants for this profession.
AI video generation — if the skill assembles visuals.
LLM pricing and token economics — what a chain run costs.
Claude Code + NotebookLM — where to send heavy research so the limits do not burn.