Claude Skills: How to Build Reusable AI Workflows

If you use Claude every day, you know the routine. You open a new chat and explain, yet again, how you like meeting notes summarised, what tone to use with clients, what your quotes look like and what to check before a post goes live. The explanation is good, the result is good, and in the next chat you start from zero.

Claude Skills solve exactly this problem. A skill is a small package of instructions, templates and sometimes scripts that Claude can pick up on its own whenever a task calls for it. You describe the process once, and from then on it works the same way in every conversation and for everyone on your team who uses it.

In this guide we cover:

• what Claude Skills are and how they are structured;
• how a skill differs from a saved prompt, a project and custom instructions;
• which business tasks are worth turning into skills;
• how to turn on ready-made skills and install one from GitHub without putting your data at risk;
• how to build your own skill step by step, and which mistakes to avoid.

What are Claude Skills?

A skill is a folder of instructions for a task you do again and again: preparing a report, editing copy to match your brand guidelines, summarising a sales call or reviewing code. At the heart of every skill is a main file called SKILL.md. It holds the skill's name, a short description of when to use it, and the instructions themselves. Next to it you can add other files: templates, examples, reference material and scripts.

What a skill is made of

• SKILL.md – the main file. It opens with two required fields: name (the skill's name) and description (what it does and when to use it). After that comes the workflow, written in plain language.
• Templates – for example, a fixed structure for meeting notes or a quote.
• Examples – a good output and a bad one, so Claude understands where the bar is.
• Reference material – a style guide, a glossary, a service price list, design rules.
• Scripts – code Claude runs when it needs an exact calculation, a file conversion or an automated check.

How Claude loads a skill step by step

The clever idea behind skills is progressive loading. At the start of a task Claude sees only the name and description of each installed skill, a few dozen words each. If the task matches a description, Claude reads the full instructions in SKILL.md. Only if those instructions point to another file does it open that file too.

This means you can install many skills without crowding the model's context window. A skill that is not relevant to the conversation costs almost nothing, so you can afford to write detailed instructions and attach long reference documents.

Where skills came from and where they work

Power users were saving instructions in separate files and pasting them into chats long before the idea had a name. In late 2025 Anthropic launched Agent Skills as an official feature and turned that habit into a consistent format. The company later published the format as an open standard, and other tools for working with AI agents have started to support it.

Today skills work in the Claude apps on the web, desktop and mobile, in the Claude Code coding agent and through the API. In other words, the same skill can serve an ordinary chat and an agent working on its own. If you are new to Claude itself, start with our introduction to Claude.

Skills vs prompts, projects and custom instructions

Claude gives you several ways to "teach" the model how you work. It helps to understand the differences so you pick the right tool for each job.

Saved prompts

A good prompt that you paste in every time does work, but it depends on people remembering to use it and pasting the latest version. With three people on a team you quickly end up with three versions. We covered how to write a strong prompt in the first place in our guide to writing AI prompts, and that is also the starting point for any skill.

Custom instructions and personal preferences

General settings such as "always reply in British English" or "I work in marketing" apply to every conversation. They are great for broad preferences, but not for a detailed process with a template and examples, because they load every time, even when they are not needed.

Projects

A project in Claude groups files and instructions around one topic, such as a specific client or a product launch. The knowledge inside a project is available only within that project.

Skills

A skill describes how to do a task, not what the conversation is about. It is available in every chat, Claude activates it on its own based on the description, and you can combine several skills in one task: for example, a data analysis skill and a presentation skill. A simple rule of thumb: if you have explained the same process to Claude three times, it is time to turn it into a skill.

Which business tasks suit Claude Skills

Skills fit any repeatable process with clear rules. These are the tasks where they save the most time.

Documents and quotes

Most business documents follow a template: quotes, meeting notes, client briefs, monthly reports. In a skill you describe what the document consists of, what it must include, how it should be worded and what a good example looks like. Claude also has built-in skills for working with Word, PowerPoint, Excel and PDF files, so you get a finished file rather than just text in the chat.

A tip for teams that work across markets: spell out in the skill which language and spelling the document uses, the date format, the currency, and whether prices are shown with or without VAT. Small details like these are exactly what the model forgets when nobody tells it.

Data analysis and reports

A skill can describe how to process a raw export from Google Analytics, your CRM or your online store: which metrics matter, how to round, which comparisons to make and which charts to build. Combine it with a document skill and you get a complete workflow: upload a file and receive a formatted report with conclusions.

Content and brand voice

An editing skill can hold your brand voice, banned words, paragraph length, headline rules and a pre-publication checklist. The same goes for SEO: you can write a fixed sequence for reviewing a page, from the title and meta description to internal links.

Replies to leads and emails

A reply to a website enquiry, a follow-up on an unanswered proposal, a sales call summary: all of these follow almost the same structure. In a skill you define what every reply must contain, what it must never promise and when to suggest a phone call instead.

Code and website review

For development teams, a skill can hold your code review rules, the order of checks before a release and troubleshooting instructions. This is especially useful in Claude Code, because the agent can run the checks itself.

How to use ready-made skills

The built-in skills, such as the ones for working with files, are on by default and Claude uses them when needed. Any other skills you add yourself. Menu names change from time to time, but in the app the process looks roughly like this.

Step 1: Turn on code execution

Skills rely on Claude's code execution environment: that is how the model reads skill files and creates new files. In Settings, under Capabilities, make sure code execution and file creation is enabled. On organisation accounts an admin may need to allow it first.

Step 2: Check what is already installed

In the Skills section you will see the list of skills, a description of each and an on/off switch. Go through the list and turn off the skills you do not use, so Claude does not accidentally pick the wrong one.

Step 3: Add a skill from the directory

Claude has a directory of ready-made skills. Choose one, read the description and add it to your list. You can remove a skill you no longer need from the same place.

Step 4: Test it in a new chat

Open a new conversation and describe the task as you normally would. In most cases Claude spots the matching skill on its own. If it does not, ask explicitly: "Use the meeting-summary skill." In Claude Code you can also call a skill directly with a slash command followed by the skill's name.

How to install a skill from GitHub safely

Many skills are published on GitHub and never appear in the official directory. Anthropic itself maintains a repository of example skills, and independent developers run collections sorted by field and profession. But a skill is not just text: it can contain scripts and instructions to contact external services. That is why you should review every skill before installing it.

Step 1: Find a skill

Search GitHub for topics such as agent-skills or claude-skills. Prefer repositories from well-known developers with a history of updates and clear documentation.

Step 2: Read every file

Open SKILL.md and every file next to it. Ask yourself: do the instructions match what the skill promises? A skill for designing presentations has no reason to request access to your inbox, send files to an external address or install extra software without explanation. If you do not read code, ask a developer to go through the scripts, or at the very least ask Claude to explain what each file does before you install it.

Step 3: Download and check the structure

In the GitHub repository, click Code and then Download ZIP. After unpacking, confirm that the folder contains a SKILL.md file and that it opens with the name and description fields.

Step 4: Upload it to Claude

In the Skills section, choose to create a new skill, then upload a file, and select the ZIP. The skill will appear in your list. Check that the name and description are correct and that the switch is on.

Privacy and data protection

If you handle personal data of people in the EU, the GDPR expects you to know where that data goes and to have a reason for processing it. This is general guidance, not legal advice. Do not put client lists, phone numbers, or health or financial details inside a skill. A skill should describe a process; the data itself arrives separately with each task, and only when there is a genuine need for it.

How to build your own skill, step by step

If no ready-made skill fits, build one. Let's walk through a simple example: a skill that turns notes or a transcript from a client meeting into a structured summary.

Step 1: Describe the task

Pick one process you keep explaining from scratch and break it into five parts:

• Input – what you give Claude: notes, a transcript, a document.
• Workflow – what to find, what to check and what to collect.
• Constraints – what must never appear in the output, such as prices that have not been agreed.
• Output format – sections, length and required fields.
• Example – what good, finished work looks like.

For instance, the description of our skill: "Summarises a client meeting using a fixed template. Use when you receive notes or a call transcript and are asked for a summary, decisions or a task list."

Step 2: Choose how to create it

There are three ways:

• In a chat with Claude. Write "I want to create a skill" and describe what you need. Claude asks clarifying questions and builds the files using a built-in skill for creating skills. This is the easiest route and requires no technical knowledge.
• From a demonstration. On some plans and operating systems you can carry out the task on your computer and let Claude learn a skill from it. Check whether this option is available on your account.
• By hand. Write SKILL.md and the supporting files yourself, zip them and upload. This gives you full control over every word.

Step 3: Create a folder

Create a folder called meeting-summary with two files inside: SKILL.md with the instructions, and template.md with the summary template.

Step 4: Write SKILL.md

The file starts with a YAML header: three dashes, the fields, and three more dashes. The two required fields are:

• name – the skill's name in lowercase letters, numbers and hyphens, up to 64 characters. It should match the folder name.
• description – what the skill does and when to use it. This is the most important field, because Claude uses it to decide whether to open the skill. Include the keywords people actually use, such as "summary", "minutes" and "action items".

After the header, write the workflow in plain sentences: read the notes, identify the participants and decisions, split action items by owner and due date, mark tasks without an owner as "unassigned", never invent details that were not said, and fill in the result using template.md. You can write the instructions in whatever language your team works in; Claude understands them well.

Step 5: Fill in the template

In template.md, lay out the structure of the summary: date and participants, meeting goal, decisions, action items (what, who, by when), open questions and the next step with the client. Because SKILL.md already points to this file, Claude opens it only when it is actually preparing a summary.

Step 6: Upload and test

Zip the folder, upload it in the Skills section and switch it on. In a new chat, paste some sample notes and deliberately include one task without an owner and one without a date, to see how Claude handles the gaps. If something does not come out the way you want, adjust SKILL.md and test again. Two or three rounds are usually enough.

Skills with scripts and reference files

For meeting summaries, instructions and a template are enough. Other tasks demand a level of precision a language model alone does not always deliver: calculations, file conversions, technical checks. That is where scripts come in.

Example: checking a web page before it goes live

Say you want a skill that reviews a new page on your website before publication. The folder might contain:

• SKILL.md – the full review sequence and the report format;
• guidelines.md – your site rules: title length, H1–H3 structure, alt text for images, internal links;
• examples.md – examples of useful review comments and unhelpful ones;
• a check script – a short piece of code that fetches the page and automatically checks meta tags, images without alt text and broken links.

Claude runs the script, gets exact results back, and then adds what a script cannot do: an assessment of the copy, the headings and the call to action. The output is a report with a short summary, a table of issues ranked by severity and a list of things that are already done well.

When a script is worth it, and when it is not

A script is worth the effort when a step has to be exact and consistent every time, or when you need to process a large file. If the task is mostly judgement and wording, good instructions and examples will work better. The more code a skill contains, the more important it is that someone technical reviews and maintains it.

Skills inside automations and agents

Skills are also available through the API and in agent-building tools, so you can use the same process outside the chat window. For example, an n8n workflow can receive an enquiry from a website form, pass it to Claude with a reply-drafting skill, and save the draft in your CRM for a salesperson to approve. We explain the difference between a chat and an agent that works on its own in our guide to AI agents for business, and how to build workflows like this in our n8n guide.

Common mistakes when building skills

A description that is too vague

A description like "helps with writing" does not tell Claude when to use the skill. State what it does, on what input and in what words people ask for it. A weak description is the most common reason a skill "doesn't work".

One skill for everything

A skill that tries to handle quotes, social posts and reports at once will be mediocre at all of them. Several small, focused skills are better. Claude knows how to combine them.

No examples

Instructions explain what to do; examples show what it looks like. Two good examples and one "not like this" improve the result more than another page of instructions.

Installing without checking

A skill from GitHub that you have not read is third-party code running with your access. Even if the repository is popular, read the files before you install it.

Never updating

Your price list changes, your proposal wording gets revised, and the skill keeps producing the old version. Decide who in the business owns each skill, and review your skills once a quarter.

How RJL Studio helps you bring AI into your workflows

At RJL Studio we build AI-powered workflows for businesses: Claude skills that know your templates, brand voice and pricing, agents that reply to leads from your website, and integrations that connect the model to your CRM, forms, WhatsApp and automation tools such as n8n, Make and Zapier. We help you decide which processes are worth turning into skills, write and test the instructions, and make sure your clients' personal data does not end up where it should not.

Want to win back hours of repetitive work every week? Read about our automation and integration services or get in touch and we will get back to you.

Frequently asked questions

Do I need to know how to code to create a skill?

No. Most skills are plain-language instructions plus a template, and you can build them in a conversation with Claude itself. Coding knowledge is only needed for skills that include scripts, and even then it is wise to have someone technical review the code.

Can I write a skill in a language other than English?

Yes. You can write the instructions, templates and examples in your team's language, and Claude works with them well. Only the skill's technical name (the name field) has to use lowercase Latin letters, numbers and hyphens.

What is the difference between a skill and a project in Claude?

A project groups knowledge about a particular topic, and what is inside it is available only within that project. A skill describes how to carry out a task, is available in every conversation, and Claude activates it on its own when a task matches its description.

Are skills from GitHub safe?

Not automatically. A skill can contain scripts and instructions to send data to external addresses. Install only from sources you trust, read every file before installing, and never put clients' personal data inside a skill.

Can I use the same skill in an automation?

Yes. Skills are also available in Claude Code and through the API, so a process you built for chat can also power an agent or an automated workflow, for example a first reply to website leads that is saved as a draft in your CRM.