Make Your Own Skill

You do not need to be able to code to teach Claude something new. If you have a task you do the same way every time — a report you build, a way you tag assets, a format you always want — you can turn it into a skill that any teammate can then run. There are three ways to make one, and the newest is the easiest: you record yourself doing the task once, out loud, and Claude writes the skill.

Anthropic’s definition

“Skills are folders of instructions, scripts, and resources that Claude loads dynamically to improve performance on specialized tasks.”

In plainer terms: a skill is a written-down playbook. Claude reads it before starting a task so you get your result, your way, without re-explaining it every time. Unlike a Project (which is always-on background knowledge) a skill only loads when it is relevant — so you can have hundreds without them getting in each other’s way.

The three ways to make one

1

๐ŸŽฌ Record it

Easiest · ~10 minutes · no writing

Hit record, do the task once while talking through it, click Done. Claude watches, then proposes the skill for you to approve.

Where: Cowork in Claude for Mac

Full walkthrough below โ†“
2

๐Ÿ’ฌ Talk it through

Easy · 15–30 minutes · a conversation

Tell Claude what the skill should do and answer its questions. Claude writes the skill file, packages it, and hands it back.

Where: anywhere you can chat with Claude

How that goes โ†“
3

๐Ÿ“ Write the file

Most control · longest · a little technical

Author a SKILL.md file yourself, optionally with reference documents and scripts, zip it and upload it.

Where: Customize > Skills

The file format โ†“

All three produce the same thing. Route 1 and 2 are how almost everyone at MM should start — route 3 is mostly useful when a skill needs bundled files, like our Brand Guardian skill which carries logos and colour codes with it.

๐ŸŽฌ Route 1 — Record a Skill, fully explained

This is the one worth learning first. Instead of describing a process in words, you demonstrate it: Claude records your screen and your voice while you do the task once, then reads that recording and writes the skill itself. You approve or reject what it proposes.

Before you start — the requirements, exactly

Step by step

  1. Open a Cowork task in Claude for Mac. Start it the way you would any task — and if the skill involves files, have the folder connected before you record.
  2. Start the recording. There are two doors to the same place: click the + in the composer and choose “Record a skill”, or go to Customize > Skills, click Add, and choose “Record your screen”. Then click Start recording.
  3. Leave your microphone on and talk. Anthropic is explicit that this is the difference between a mediocre skill and a good one: “Talk through what you’re doing while you record. Narration gives Claude context it can’t get from your screen alone, like why you skip a step or how you choose between two options.” You can mute or switch input with the microphone control if you need to.
  4. Do the task, start to finish, the correct way. A recording runs for about 10 minutes, which is plenty for one task and not enough for three. Claude is learning from what it sees, so a fumbled or abandoned step teaches it the fumble.
  5. Click Done, and let Claude review. Claude watches the recording back and works out what the repeatable process actually was.
  6. Approve what it proposes. Claude comes back with a proposal card. If it decided this is something new you get Save or Dismiss. If it recognised the task as an existing skill it offers Update or Dismiss instead — which is how you improve a skill over time rather than piling up near-duplicates.
  7. Test it before you trust it. Open a fresh task and ask for the thing in your own words, without naming the skill. If Claude picks it up, the description is doing its job. If it does not, the description needs work — see the description section below.

โš ๏ธ Read this before you press record

Anthropic’s warning, word for word: “Don’t type passwords or secrets, or display sensitive information or private conversations while recording. Everything on your screen is captured for the length of the session, along with anything you say.”

What that means for us specifically: before you hit record, close anything showing customer names, addresses, emails or order details, vendor contracts or pricing, financials or payroll, and Slack DMs or personal email. Do not sign in to anything on camera. If a step needs a password, pause and do that part off-recording. If you record something you should not have, dismiss the proposal and delete the task.

๐Ÿ”’ What is actually kept afterwards

Anthropic: “The video and audio from your recording aren’t retained.” Claude reviews the recording in order to build the skill, and what survives is a set of screenshots from the session. You can look at them yourself by expanding the Recorded demonstration step inside the task. Those screenshots live in that Cowork task — delete the task and they go with it.

So your narration is used to build the skill and then discarded; the screenshots are the part that sticks around, which is another reason to keep sensitive things off screen.

What makes a recording produce a genuinely good skill

One task, not a tour. “How I tag a new product shoot in Canto” is a skill. “How I do my job” is not. If the task has two halves that you sometimes do separately, record two skills.
Rehearse once, then record. Ten minutes is generous for a clean run and tight for a messy one. Know where your files are before you start.
Narrate the judgement, not the clicks. Claude can see that you clicked the third option. It cannot see that you chose it because the second one is only for wholesale. Say the why out loud.
Use a real example, not a dummy one. Real inputs surface the real edge cases — as long as the real example is not sensitive.
Say what “good” looks like at the end. Finish by describing the finished result and how you would know it was wrong. That is the part people always leave out.
Then improve it by recording again. Claude offering Update on an existing skill is a feature, not a mistake. The second recording of the same task is usually the one that makes the skill reliable.

๐Ÿ’ฌ Route 2 — Talk it through

This works everywhere, including Windows, and it is the right route when the skill is about knowledge and standards rather than a sequence of clicks — a tone of voice, a report format, a checklist.

  1. Open a new chat and say what you want. Something as plain as “I want to create a skill for [your task]” or “I need a skill that knows how to [your process]”. Attach anything that shows the standard — templates, past examples, brand guidelines, a spreadsheet.
  2. Answer Claude’s questions properly. It will ask things like when you would use this and what makes the output good. Anthropic’s guidance is to “provide enough detail that someone capable but unfamiliar could follow your approach” — write for a competent new hire, not for yourself.
  3. Claude builds it. It writes the SKILL.md, organises whatever you attached, generates any code the process needs, and packages the whole thing into a single skill file.
  4. Turn it on and test it. Your skills live under Customize > Skills (also reachable as Settings > Capabilities > Skills). Toggle it on, then ask for a task it should cover. When Claude picks it up you will see “Using [skill name]” in its thinking.
  5. Refine by asking. Tell Claude what was off and have it update the skill. You do not edit the file by hand unless you want to.

๐Ÿ“ Route 3 — Write the file yourself

Worth understanding even if you never do it, because it explains why the other two routes ask what they ask. A skill is just a folder with a markdown file in it.

Structure — the folder must be the root of the zip, not the loose files
my-skill.zip
โ””โ”€โ”€ my-skill/
    โ”œโ”€โ”€ SKILL.md          โ† required
    โ””โ”€โ”€ resources/        โ† optional: templates, reference docs, scripts

The SKILL.md opens with a short block of settings, then the instructions in plain markdown. Two settings are required:

nameA human-friendly name. Max 64 characters.
descriptionWhen and why to use this skill. Max 200 characters — and Anthropic calls this field critical, because it is the only thing Claude reads when deciding whether to reach for your skill.

Optionally you can add reference files, executable scripts (Python or Node, with pandas, numpy and matplotlib available), and a dependencies line naming packages a script needs. Then zip it, go to Customize > Skills, click + Create skill, choose Upload a skill, and upload the zip.

The description is the whole ballgame

Nine times out of ten, a skill that “does not work” is a skill that never got triggered. Claude decides whether to load your skill by reading its description, so a vague description means a skill that sits there unused. However you build it, this is the field to fuss over.

โœ— Too vague to trigger

“Helps with Canto stuff.”

Claude has no idea what a request for this looks like.

โœ“ Triggers reliably

“Tag, keyword and album new Melinda Maria product photography in Canto to MM naming conventions. Use when uploading or organising shoot assets.”

Names the task, the system, and the moment it applies.

Getting your skill to the rest of the team

A skill you make is private to your own account by default. It does not appear for anyone else until it is deliberately shared, and there are three routes — only one of which is open to everyone today:

Publish it for the whole company

Only an organization owner can add or remove company-wide skills, under Organization settings > Skills โ†’ + Add. Once added it shows up in everybody’s Customize > Skills, switched on by default, and any individual can toggle it off for themselves.

At MM that means: send it to David and he publishes it.

Share with specific colleagues

Members can share a skill with named colleagues, and can publish to an organization directory that anyone can install from — but both of those settings are off by default and have to be switched on by an owner first.

Worth knowing that org-directory publishing has no approval step once enabled, which is why it stays off for now.

Just hand over the file

Any skill can be exported as a .zip and sent to someone, who uploads it under Customize > Skills. Low-tech, works today, no admin involvement.

Fine for one or two people. For anything the whole team should have, use route one.

Things at MM that are worth recording

If you are looking for a first skill to record, these are real MM processes already identified as repetitive enough to be worth capturing — two of them are on the projects board waiting for someone to record them:

Where this came from

Everything above about how the feature behaves is taken from Anthropic’s own documentation rather than summarised second-hand, because the details differ by plan and platform:

How to create custom skills โ†— · What are skills? โ†— · Use skills in Claude โ†— · Provision and manage skills for your organization โ†— · Create a skill through conversation โ†—

Plan and platform limits change. If something here does not match what you see on your screen, tell David and this page gets corrected.

โ† Back to all skills