How to Record Repeatable Skills in Claude

2026-08-27

Key Takeaways

Claude turns a screen recording into a reusable skill. Open the Claude desktop app, start a Cowork task, choose “Record a skill” from the + menu, perform the workflow while narrating it, and Claude drafts the skill for you.
Anthropic launched the feature on 21 July 2026. Its help documentation lists it for Pro, Max and Team plans in Claude for Mac, with recordings capped at roughly 10 minutes.
A skill is a folder containing a SKILL.md file with YAML frontmatter, plus optional scripts, templates and reference documents.
Only the name and description stay in context permanently, costing about 100 tokens. The instruction body loads when triggered, and bundled files load only when read.
You can also write skills by hand and upload them as a ZIP at Customize > Skills on claude.ai, or drop a folder into ~/.claude/skills/ for Claude Code.
Skills do not sync between surfaces. A skill uploaded to claude.ai is not available through the API, and Claude Code skills live on your filesystem separately.
Anything visible on screen during recording can end up in the skill, so use sample data and hide credentials before you press record.

Claude Code logo.

Claude records a repeatable skill by watching you do the work once. In the Claude desktop app, start a Cowork task, open the + menu, select “Record a skill”, then carry out the process while explaining your decisions aloud. When you stop, Claude converts the demonstration into a saved skill it can run again without the walkthrough.

Anthropic described the mechanic in one line at launch: “Record your screen while you do a task, talk through it as you go, and Claude turns it into a skill it can run again.” That is the whole feature. The demonstration replaces the prompt — this is the only way to build a Claude skill that requires no writing at all.

What a Claude Skill Actually Is

Before recording anything, it helps to know what the recording produces. A skill is a folder. Inside it sits a SKILL.md file with YAML frontmatter at the top and instructions beneath, alongside any supporting files the task needs.

skill-name/
├── SKILL.md
├── REFERENCE.md
├── scripts/
│   └── process.py
└── templates/
    └── report_template.docx

The frontmatter carries two required fields. The name identifies the skill, capped at 64 characters. The description tells Claude what the skill does and when to use it — the platform documentation allows up to 1,024 characters, while the claude.ai upload path asks for 200 or fewer. That description is the trigger. Claude matches incoming requests against it, so a vague description produces a skill that never fires.

Why Skills Cost Almost Nothing Until Used

Skills load in three stages, which is what makes it practical to keep dozens of them enabled at once.

Level	When it loads	Token cost	Content
Metadata	Always, at startup	About 100 tokens	Name and description only
Instructions	When the skill is triggered	Usually under 5,000 tokens	The SKILL.md body
Resources and code	Only when referenced or run	None until accessed	Extra markdown, scripts, templates, schemas

Scripts are the efficient part. Claude executes them through bash, and only the output enters the conversation — the code itself never consumes context. That makes deterministic steps such as file conversion or validation far cheaper than asking a model to reason through them each time.

Recording a Skill: Step by Step

1. Check You Have the Right Setup

Update the Claude desktop app first. Anthropic’s help documentation lists recording as available on Pro, Max and Team plans in Claude for Mac, and states it is unavailable on Windows, on the Free plan and on Enterprise. Recordings run to roughly 10 minutes. Some coverage at launch reported Windows availability too, so check the + menu in your own build rather than trusting a secondhand list. Costs across the plans are laid out in our breakdown of Claude Cowork pricing.

2. Prepare the Screen Before You Press Record

This is the step people skip and regret. Everything visible becomes evidence for the model. Close notification banners, sign out of unrelated accounts, hide API keys and tokens, and swap real customer data for a sample file. A skill built on a live production spreadsheet can carry names, account numbers or internal URLs into a document you later share with colleagues.

3. Start the Recording From a Cowork Task

Open a Cowork task, click the + button in the composer, and choose “Record a skill”. The recording captures screen activity, mouse movement, clicks and keyboard input, along with your spoken audio.

4. Narrate the Rules, Not Just the Motions

The spoken track is where the reasoning lives. Clicking through a workflow silently teaches Claude a sequence of movements. Saying “I always check the invoice date first, and if it falls outside the quarter I move the row to the deferred tab” teaches it a rule. Explain why you choose a file, why you reject a row, what you do when a value is missing, and what counts as finished.

5. Stop, Then Review What Claude Wrote

Claude converts the demonstration into a skill proposal and saves it to your skill library. Read it before enabling it. Check three things in particular: whether the description states both the purpose and the trigger conditions, whether any identifiers or credentials survived from your screen, and whether an edge case you handled by instinct made it into the written steps.

6. Test on Data It Has Never Seen

Run the skill against a fresh file rather than the one you demonstrated with. A recorded skill can quietly encode the specifics of a single example — a column position, a filename pattern, a date format. Testing on new input exposes that immediately.

The Manual Route: Writing SKILL.md Yourself

Recording suits visual, repetitive desktop work. For anything with precise rules, formatting standards or scripts, writing the file by hand gives you tighter control. The minimum viable skill looks like this:

---
name: quarterly-invoice-check
description: Validates quarterly invoice spreadsheets against the finance checklist. Use when the user uploads an invoice export or asks to reconcile a quarter.
---

# Quarterly Invoice Check

## Instructions
1. Open the uploaded spreadsheet and confirm columns match the expected schema.
2. Flag rows where the invoice date falls outside the stated quarter.
3. Move flagged rows to a sheet named "Deferred".
4. Produce a one-paragraph summary of totals and exceptions.

## Examples
[Include one worked example with input and expected output]

Package the folder as a ZIP with the skill folder at the root — not nested inside another directory — then upload it at Customize > Skills on claude.ai, reachable at claude.ai/customize/skills. Anthropic also ships a skill-creator skill that interviews you and writes the file, which is a middle path between recording and hand-authoring. If building automated helpers is new territory, our beginner’s guide to creating AI agents without coding covers the same ground from the ground up.

Where Skills Live on Each Surface

This trips people up constantly: skills do not follow you between products.

Surface	How to add a skill	Who can use it	Network access
claude.ai and desktop	Upload a ZIP at Customize > Skills, or record one	You only, no org-wide sharing	Varies by user and admin settings
Claude Code	Folder in ~/.claude/skills/ or .claude/skills/	Personal or per-project	Full
Claude API	Upload via the /v1/skills endpoints	Everyone in the workspace	None, and no runtime package installs
AWS and Microsoft Foundry	Skills API, same as the Claude API	Workspace	Same limits as the API

Four prebuilt document skills — pptx, xlsx, docx and pdf — are available on the API, on claude.ai and through the cloud platforms, but not inside Claude Code. Claude Code instead bundles the open-source Claude API skill. Developers working there will find more context in our guide to Claude Code as an agentic terminal assistant.

Security Rules Worth Following

A skill is executable instruction. Treat one the way you would treat a script someone emailed you.

Install skills only from sources you trust, and read every file before enabling — SKILL.md, scripts, templates, the lot. Warning signs include network calls the skill has no reason to make, file access outside the stated scope, and any instruction to fetch content from an external URL at runtime. Never hardcode API keys or passwords into a skill file. Enterprise customers get skill content scanning, though it does not cover uploads made through the API or Console.

One more detail worth knowing for regulated environments: Agent Skills are not covered by Zero Data Retention. Skill definitions and execution data are retained under Anthropic’s standard policy, and audit logging is available for Skills API operations.

Writing Skills That Actually Trigger

Most failed skills fail at the description, not the instructions. Claude only sees the name and description when deciding whether a skill is relevant, so the description has to carry both halves of the job: what the skill does and when it applies.

Compare “Handles reports” with “Builds the weekly sales report from a Salesforce CSV export. Use when the user uploads a Salesforce export or asks for the weekly sales summary.” The second version names the input, the output and the trigger phrase. That is the difference between a skill that fires and one that sits unused.

Keep the SKILL.md body under roughly 5,000 tokens and push detail into separate files that Claude reads only when needed. Bundle a worked example. Use scripts for anything that must produce the same result every time. The same discipline that makes prompts reliable makes skills reliable, and several of the habits in our tips for getting the most out of Claude Code apply directly here.

When Recording Beats Writing, and When It Doesn’t

Record when the task is visual and hard to describe — moving files between folders, cleaning a spreadsheet layout, navigating a web tool with an awkward interface, assembling a recurring report from three sources. Watching is faster than writing 40 steps.

Write by hand when the task depends on rules rather than motions — style guidelines, validation logic, formatting standards, anything that needs a script for determinism. And write by hand when the skill will be shared, because a hand-authored file has no risk of carrying stray screen content.

The strongest approach combines both. Record once to get a first draft, then open the generated skill and edit it into shape: tighten the description, remove the incidental, add the edge case you forgot to narrate, and attach a script for the fiddly part.

If you are interested in this topic, we suggest you check our articles:

Sources: Claude Platform Docs — Agent Skills, Anthropic Help Center, Digital Applied, Anthropic Skills Repository, Cyber Security News

Written by Alius Noreika