Every Monday, you open a chat with Claude and paste the same block of instructions. The tone to keep, the shape of the deliverable, the three rules the model always forgets. Claude skills exist to kill that habit.
A skill is a folder of plain-language instructions that Claude loads on its own when the task matches. You describe your method once, in a file called SKILL.md. Claude reapplies it from any conversation, without you repeating anything.
Below: how a skill gets triggered, how to write one without coding, how to repair a skill that never fires, and where skills stop being useful. No programming background needed to follow along.
Skill, prompt, project, connector: who does what
Four mechanisms look alike from a distance and do different jobs. Mixing them up is the number one reason people end up with useless skills.
| Mechanism | What it does | Lifespan |
|---|---|---|
| Prompt | One instruction typed into the chat | Forgotten in the next chat |
| Custom instructions | Background behaviour, a general tone | Applied everywhere, all the time |
| Project | A space that keeps the context of one file of work | Limited to that project |
| Skill | A full method, with its steps and its output format | Permanent, fires when the task fits |
| MCP connector | Access to an outside tool (mail, CRM, browser) | Permanent, as long as the connection holds |
A skill teaches Claude how you work. A connector tells it where your data lives. They pair well, and neither one replaces the other.
MCP stands for Model Context Protocol, the open format published by Anthropic to plug a model into external tools. A skill only handles text and local files. It will not fetch anything from the internet by itself.
How Claude decides to activate a skill
The mechanism runs in three beats. Understanding it removes half of the beginner mistakes.
- Discovery: when a session starts, Claude reads only your skills' metadata, the name and the description. That stays cheap.
- Activation: when you write your request, Claude compares it to those descriptions and picks the matching skill itself. You can also call it by name.
- Execution: only then does Claude open the full file, load the document templates or scripts it needs, and follow the steps.
The practical consequence is simple. Your description matters as much as your instructions, because it is the only thing Claude looks at when deciding. Perfect instructions behind a vague description give you a skill that sleeps.
A skill never fires on its own. It needs a message from you inside a conversation. This is not a scheduled job that runs at 7 a.m. while you sleep.
If you want truly hands-off execution, triggered by an incoming email or a fixed hour, you need an automation tool alongside Claude. That is what n8n does, and we cover that approach in our course on building automations with n8n.
Anatomy of a SKILL.md file
A skill fits in one Markdown text file, split in two parts. At the top sits the frontmatter, a metadata block wrapped in two lines of three dashes. Underneath come the instructions.
---
name: client-meeting-notes
description: Writes a client meeting report in the agency format.
Use when the user provides meeting notes, a call transcript,
or asks for a recap after a prospect meeting.
---
# Client meeting report
## Steps
1. Identify the client name, the date and the attendees.
2. Summarise the stated need in 5 lines maximum.
3. List the agreed actions, each with an owner and a due date.
4. Write the follow-up email, direct tone, 150 words maximum.
## Output format
- A table of actions (action / owner / due date)
- The follow-up email last, ready to copy
## Edge cases
- If no due date was given, write "to be confirmed".
- Never invent an amount or a contractual commitment.
The name field is lowercase with dashes. The description field tells Claude when to wake up, using the words you actually type when the situation comes up.
For heavier skills, you can build a folder around that file. A templates/ subfolder for your document models, references/ for your business rules, scripts/ for executable code. The main file points to those resources, and Claude loads them only when needed.
Writing the instructions: three levels of control
Not every skill needs the same rigour. Set the level of detail against the cost of a mistake, not against your urge to do things thoroughly.
- Loose: brainstorming, rewriting, headline ideas. A few lines are enough, let Claude interpret.
- Medium: meeting reports, weekly updates, summaries. A known format, with acceptable variations.
- Tight: accounting, legal work, compliance. Every step spelled out, every edge case anticipated, plus two or three examples of a correct output pasted into the file.
A common mistake turns the skill into a lecture. Claude already knows how sales tax works. What it does not know is your processing order and your formatting.
Three ways to create your first skill
1. Let Claude write it for you
The fastest route is a plain conversation. Describe the task you repeat, paste two or three past outputs as examples, and ask Claude to build the skill. Anthropic maintains an official skill for exactly this, the skill-creator, which asks the questions on your behalf.
Always read the generated file before you save it. That is where you catch a description that is too vague, or a step you would never have phrased that way.
2. Write the file by hand and upload it
Open a text editor, write the frontmatter and the instructions, save as .md. On Windows, turn off hidden file extensions so you can check the file is not actually named skill.md.txt.
You then add it from Claude's settings, under the capabilities and skills sections. The exact wording changes between interface versions. Remember to allow code execution, otherwise skills that handle files will not work.
3. Go through Claude Code, to version them
In Claude Code, skills live as files in the project's .claude/skills/ folder. Git tracks them like any other piece of code, and the whole team inherits the same context.
/plugin marketplace add anthropics/claude-plugins-official
/plugin install skill-creator@claude-plugins-officialIf the first command errors out, the repository name has changed. Type /plugin alone to list the available sources. Anthropic also publishes example skills on GitHub, worth reading before you write your own.
A fourth route exists in Claude Cowork, the desktop app. You record your screen while doing the task, explaining your choices out loud, and Claude turns that into a skill. We walk through that environment in our Claude Cowork course.
The recording captures everything on screen, not just the window that matters. Close your inbox, your chat apps and your client files before you hit record.
One nuance about sharing. Sending a skill to a colleague in one click depends on the team plans. On an individual plan, you hand over the file and the other person installs it themselves. Check what your subscription covers on Anthropic's pricing page.
A skill that never fires: the repair job
This is problem number one. You write a message that should have triggered your skill, and Claude answers as if it did not exist. The culprit is almost always the description.
Rewrite it with the phrasings you genuinely use. Not "financial document processing" but "when I mention an invoice, an expense claim or a bank statement". The trigger words should sound like your real sentences.
Then test on a real case, and change one thing at a time. Either the description, or one workflow step, never both. Otherwise you will not know which of the two edits produced the effect you observed.
In Claude Code, the skill-creator can launch several sub-agents in parallel, half of them using the skill and half without, then compare the results. That test burns tokens (the billing units of a language model), so save it for the skills you run daily.
Get into the habit of rerunning that test when a new model ships. A skill written to compensate for a weakness can start working against you once the next model handles that area well on its own. Generation jumps, like the one brought by Claude Opus 5, sometimes change the default behaviour across a whole domain.
Chaining several skills on one process
One skill does one thing. To automate a whole process, you create several and call them in order from a main skill.
Take a weekly editorial routine. A first skill collects the topics you have been tracking. A second scores them by interest. A third drafts the posts from the best-scored ones. A fourth calls the other three and compiles everything into a single file.
The golden rule fits in one sentence. Break the manual process into steps, validate each step on its own, and only assemble afterwards. A skill built on a shaky step just automates the flaw, and you will spend longer hunting it down than you spent writing the thing.
Which tasks deserve a skill
A good candidate ticks three boxes. You do it often, always the same way, and the expected result has a recognisable shape. Meeting notes turned into a report, a weekly update, a proposal built from rough notes, a code review against your internal conventions, unit test generation.
A rough calculation to decide. A 40-minute task done 4 times a month costs you 32 hours a year. Writing the skill and fixing it usually takes one to three hours. So it pays for itself by the second month, even if the skill only saves half the time.
Two kinds of tasks resist. The ones you do twice a year, where writing the skill costs more than the task itself. And the ones whose structure changes every time, because a skill earns its keep through repetition. On bespoke work it becomes a straitjacket.
There is a second, less obvious use. A skill can force Claude to think before it executes, in an area where it rushes. An interface generated with no guidance often comes out grey, with whatever typeface was handy, while a design skill makes it settle those questions first. That idea runs through our Claude Design course.
The limits and risks worth knowing
Dozens of community skills float around online. Many are old prompts reformatted, with nothing new inside. Sorting through them takes time, and writing your own is often faster than hunting for the right one.
A skill can carry malicious instructions or code, invisible unless you read the file. Always open the contents before installing, and activate nothing that comes from a source you cannot identify.
The other way round, never put a secret in a skill. No password, no API key, no client personal data in the instructions or the metadata. The file is built to be shared, so treat it that way and assume it will leave your machine one day.
One last point, the one most often missed. A skill multiplies one person, it does not replace a team. It removes the formatting, the copying, the repeating of rules. It does not make the judgement call, and it does not take responsibility for proofreading a document that commits you.
Frequently asked questions about Claude skills
Do you need to code to create a Claude skill?
No. A skill is a text file written in Markdown, readable by a human, with no programming language involved. You can even have Claude write it by describing the task and showing two or three examples of your past work. Coding only becomes useful if you want to attach scripts or track your skills in Git.
What is the difference between a Claude skill and an MCP connector?
A skill encodes a working method: the steps, the rules, the output format. An MCP connector gives Claude access to an outside tool, an inbox or a CRM for example. The first teaches it how to work, the second where to find the material. Both run together without interfering.
Can a skill replace an AI agent or an automation tool?
No, because a skill never starts by itself. It needs your message in a conversation to fire. For execution triggered by an incoming email or a fixed hour, you need an automation tool such as n8n. For most of the tasks you already do at your screen, a skill is enough.
My skill does not activate, what should I do?
Rewrite its description, since that is the only thing Claude reads when deciding. Replace generic wording with the words you actually type when the situation comes up. Then test on a real case, changing one element at a time, either the description or the instructions, never both together.
How many repetitions make a skill worth it?
Writing a skill and fixing it takes one to three hours. A 40-minute task repeated 4 times a month adds up to 32 hours a year, so the investment pays back by the second month even with a partial gain. Below two occurrences a year, keep a prompt you adapt by hand.
