Dev48
Language
  • About
  • Services
  • Industries
  • Technologies
  • Articles
  • Contacts
Book a call
    Home/Articles/Claude cowork skills how they work and what to build
Dev48
info@dev48.org+375 29 320-40-96WhatsApp@isky_by

© 2026 · All rights reserved.

Claude Cowork skills: How they work and what to build

Source: Hubspot•November 19, 2026 at 02:26 PM

Claude Cowork skills: How they work and what to build

If you’ve yet to understand Claude Cowork skills, let me help you out: Claude Cowork is a mode that can access your desktop files and complete tasks for you. A skill is a set of instructions that tell Claude exactly how to perform a specific workflow. When combined, the two can boost your…

If you’ve yet to understand Claude Cowork skills, let me help you out: Claude Cowork is a mode that can access your desktop files and complete tasks for you. A skill is a set of instructions that tell Claude exactly how to perform a specific workflow. When combined, the two can boost your productivity in a way that Claude Chat alone cannot.

Claude skills aren’t limited to one Claude surface or tool. Skills work in Claude Chat, Claude Code, and Claude Cowork. So “Claude Cowork skills” are actually a feature across the AI, not just a single mode.

This is even truer since Anthropic announced that Claude Cowork, Chat, Artifacts, and Design are merging into a single unified experience, starting in October 2026. Skills, projects, connectors, and tasks all carry over, but users no longer have to dance between tabs. Claude simply decides which tools to use based on what you ask.

These changes are meant to reduce confusion, but it’s natural to still have questions. We’ll clear things up as they stand now, share what Claude Cowork skills are, and what you should build to get the most out of them.

Note: These changes are rolling out to Claude’s Pro and Max plans first, on the web, desktop, and mobile over the coming weeks, with plans to extend them to the free and team tiers later.

Table of Contents

  • What are Claude Cowork skills?
  • Claude Skills vs. Projects vs. Instructions vs. Memory
  • How Claude Cowork Skills Work
  • How to Create Your First Claude Cowork Skill
  • Claude Cowork Skills: Best Uses for Revenue Teams
  • Claude Cowork Best Practices for Safety and Reliability
  • Connect Claude to HubSpot
  • Frequently Asked Questions About Claude Cowork Skills
  • All it takes is one skill to start.

A Claude Cowork Skill is a reusable instruction folder (with optional supporting templates, scripts, or reference documents attached) that Claude loads when a task needs it, and ignores when it doesn’t. You can use Claude Cowork skills across every other Claude surface, so truly, any Claude skill is a Claude Cowork skill and vice versa.

If you want to get into the nuts and bolts of it, a Claude skill is technically a folder that can contain the following:

  • Instructions: the step-by-step procedure and the rules Claude should follow.
  • Templates and reference files: brand guidelines, sample outputs, or layouts that show the standard to match.
  • Scripts: code for the deterministic parts, such as reformatting data or generating a file.

What makes a skill different from a detailed prompt you paste every time is:

  • It’s portable. One skill works in Chat, Cowork, and Claude Code; you don’t rewrite it for each surface. You can also use Claude skills in different AI or LLM tools like ChatGPT.
  • It’s consistent. Claude follows the same steps, in the same order, every time. No drift, no improvising.
  • It’s smart about context. Claude only loads what it needs, which keeps the conversation focused and efficient.

Thanks to Cowork and the recent merger, a skill can now also access internal documentation and do things like reading and writing to folders on your computer, running multi-step tasks autonomously, coordinating sub-agents, and producing finished files like spreadsheets and presentations, not just responding in text like it would with Chat alone.

When you need Claude to repeat a multi-step process the same way every time (i.e., formatting a weekly report, triaging leads, summarizing support tickets), that’s where a skill comes in handy.

For example, when I’m using Claude for fact-checking assistance, I often prompt it with detailed instructions on what to verify in the article, how to verify it, and what to include in its citations so that I can double-check its work. Instead of repeating that prompt every single time I’m done with an article, I have a “blog-fact-checker” skill that contains my detailed instructions. Now, instead of 500 words, I simply type two words, “Fact check,” and attach my draft via Google Docs. Easy.

Note: Record a Skill is a Cowork feature still exclusive to Claude Cowork on Mac desktops with Pro, Max, and Team plans, but we’ll get into that later.

Where to Access Claude Skills

Claude automatically selects and uses skills depending on the context. You don’t have to tell it when to use a skill or which one to invoke, though you can do so in a few ways:

1. When creating a skill, you can include explicit wording in the skill description on how Claude can know when to invoke the skill (Anthropic recommends writing in third person). Example: “Use this skill whenever the user says ‘generate an image,’ ‘make an infographic,’ or ‘create an image.’”

2. Add the instructions to your prompt. Example: “Use your infographic skill to create a bar chart showing the monthly sales comparison.”

Claude Skills vs. Projects vs. Instructions vs. Memory

As a marketer, I’m often overwhelmed by the sheer number of features that Claude offers (and new ones get released seemingly every week!). So, here’s a quick demystifier on the difference between Claude skills, projects, instructions, and memory.

  • Memory is what Claude learns from your chats (either automatically or because you told it to remember). There is also project-specific memory, where Claude will remember info relevant only to that specific project.
  • Skills are repeatable instructions that can be invoked explicitly (you tell Claude to use it) or implicitly (Claude automatically figures out which skills are needed for that task). Skills are usable across your projects and tasks.
  • Projects are folders that organize related tasks and chats by topic, each holding its own files and context.
  • Instructions are broad directives that shape how Claude works, either account-wide by default or project-specific.

Which one you use depends on what problem you’re running into:

  • If you keep retyping the same prompt, create a skill.
  • If you want your files and campaign context in one place, create a project.
  • If you need Claude to remember an important detail across chats, add it to memory.
  • If you need Claude to behave a certain way every time, write instructions.

As Alex Banks puts it: “Skills are the tools in the workshop. Projects are the workshop itself.”

These four can also be combined: A single project can carry its own instructions and project-specific memory while you invoke skills inside it.

Claude plugins bundle skills, MCP connectors, commands, and agents into one package.

How Claude Cowork Skills Work

Adding a skill doesn’t “retrain” or “fine-tune” Claude, and it doesn’t change how the model thinks. A skill is simply a set of task-specific instructions Claude reads at the moment you need them, like handing a colleague an SOP and asking them to follow it.

Claude reads the skill using progressive disclosure: It brings a skill into the session piece by piece as needed, rather than all at once. Progressive disclosure works in three stages:

  • Metadata (always on): At startup, Claude loads only each skill’s name and description. That’s how it knows which skills exist and when to reach for one, and it’s why you can keep dozens installed without slowing anything down.
  • Instructions (loaded when triggered): Once your request lines up with a skill’s description, Claude pulls in the full instruction set, including the steps, rules, and formatting it should follow.
  • Resources and scripts (loaded on demand): The templates, reference files, and code inside the skill folder stay put until a task calls for them, so they cost nothing until used. When Claude runs a script, only the result enters the chat, not the code itself.

The benefit of this is that a skill can bundle large brand guidelines, datasets, or utilities without slowing down every conversation, because none of it loads until Claude needs it.

How to Create Your First Claude Cowork Skill

This article covers two ways to create a Claude Cowork skill:

  • Prompt Claude in conversation to build a skill based on your specifications.
  • Use “Record a skill” in Claude for Mac desktop app and let Claude build a skill based on the video recording.

If you already have a skill file, you can also upload it as a ZIP from Customize > Skills.

When I first learned about Claude Cowork skills, I had a hard time deciding what exactly I should turn into a skill. If you’re in the same boat, I recommend starting with any saved prompts or those you repeatedly type. Comb through your old chats and Claude projects and see what prompts or instructions you frequently use. Or if you have a prompt library in a spreadsheet or doc, mine that for ideas of what you can turn into a skill.

You can also spend one day observing what tasks you repeatedly carry out but have never documented. Capture these workflows by using Claude’s “Record a skill.”

When to Use “Record a Skill”

Claude for Mac’s “Record a skill” feature is perfect for multi-step workflows you do all the time but have never formally documented. Anything where you can record a screenshare video while verbally walking through each step in under ten minutes is an ideal candidate. Some examples:

  • Webinar lead routing: Sort an attendee export into sales, customer success, and marketing follow-up (the skill I’ll record below).
  • Security questionnaire first pass: Match a prospect’s questions to your approved answers and flag any without one.
  • Refund request review: Compare each request against your refund policy and mark it Approve or Escalate.

How to Use “Record a Skill”

Below, I’ll walk you through the steps to use the “Record a skill” feature to make your first Claude Cowork skill. To demonstrate, I created a test skill for routing webinar leads for a fictional company using fake contacts.

Important steps before you get started:

  • Check that you have the Pro, Max, or Team plan and are using Cowork in the Claude for Mac or Windows desktop app. “Record a skill” is not available on Free or Enterprise plans, and it can’t be accessed via Claude Chat (if you don’t yet have the unified Cowork/Chat experience).
  • Check that Code execution and file creation is turned on under Settings > Capabilities. If you’re on an Enterprise account, an Owner must enable Code execution and file creation as well as Skills in Organization settings > Plugins & skills.
  • Hide any passwords, secrets, private conversations, or other sensitive information before you hit record. During the session, “Record a skill” captures everything on your screen and what you say aloud, so Anthropic warns against displaying sensitive information.
  • Make sure the process you’re about to record does not take more than 10 minutes — the recording will cut off at around 10 minutes.

1. In a Claude desktop app session, click + and select Record a skill from the dropdown menu.

2. You’ll get a pop-up. Click the Start recording button.

3. Start showing the workflow you want to turn into a skill, and be sure to narrate each step that you’re taking. Think of it like filming a Loom for a coworker — you need to give all the context so Claude knows why you did each thing. You’ll know Claude is recording when a coral-colored halo surrounds your screen, and an overlay appears at the bottom showing “Capturing” and a timer.

4. When you’re done showing the workflow, click the Done button.

5. Once you click Done, Claude will automatically open a new session with a prompt that says “I recorded a demonstration of a desktop workflow for you to learn from” and start building the skill. You do not have to do anything but wait for it to finish.

6. When Claude is finished and delivers the skill, click Content to expand the instructions and read the skill to ensure it’s what you want.

7. After reviewing the contents, if you’re happy with it, click the Save button. If you want to make changes, ask Claude to make updates to the skill within the chat. It will then deliver an updated skill, and you must click the Update button to save the changes.

8. After you save your new skill, check to make sure it appears in Customize > Skills. Now, it’s ready to use! Click + and Add folder to give Cowork access to the folder with the files it needs to read and modify for this task. I recommend testing the Claude Cowork skill on dummy data first to ensure it works properly before trying it on a real task.

I tested my new webinar-lead-routing skill on fake contacts, and it worked exactly as I’d described it in the recording.

Claude Cowork Skills: Best Uses for Revenue Teams

Marketing Team Use Cases

Proving month-end marketing wins to leadership can be a challenge when you’ve got so many reports and dashboards to pull from. A performance-memo skill turns a folder of channel exports into a one-page summary in your preferred reporting format.

What goes in the skill:

  • Instructions: Compare each metric to the previous month and to the plan, flag any swing over 10%, and write “Cause unknown” rather than guess
  • Reference file: Your metric definitions and quarterly targets
  • Template: Headline, wins, risks, and next month’s focus

Example input: “September exports: Google Ads, LinkedIn Ads, GA4, webinar platform. Plan: 1,200 registrations, $85 cost per lead.”

Example output:

  • Win: Google Ads cost per lead came in at $71, about 16% under plan.
  • Risk: Registrations reached 930, about 23% under plan. Cause unknown, though the LinkedIn export shows the promo campaign paused Sept. 14.

Two more marketing use cases:

  • Competitor launch teardown: Save a competitor’s announcement and pricing page, and get a battlecard-format brief on what changed and which of your claims it affects.
  • Win/loss interview synthesis: Upload a quarter’s anonymized win/loss interview transcripts and get the top reasons deals were won or lost, with supporting quotes and the messaging changes each insight suggests.

Sales Team Use Cases

When a prospect sends contracts back with redlines, it can take days for legal to review, which can stall the deal. A redline-triage skill sorts a prospect’s changes against your approved fallback positions before the file reaches legal — so the team doesn’t waste time reviewing cases already outlined by your playbook.

What goes in the skill:

  • Instructions: Label each changed clause Accept, Counter, or Legal review; route anything the playbook doesn’t cover to Legal review; never draft final contract language
  • Reference file: Your playbook, with fallback positions by clause
  • Template: A clause-by-clause list, then a two-line rep summary

Example input: “Copperline Foods returned our MSA. Payment terms changed from Net 30 to Net 60, liability cap raised to 2x annual fees, new termination-for-convenience clause.”

Example output:

  • Payment terms: Counter. Playbook allows up to Net 45 without approval.
  • Liability cap: Accept. Within approved range.
  • Termination for convenience: Legal review. No playbook position exists.
  • Rep summary: “Two of three changes can be resolved today. Hold the termination clause for legal.”

Two more sales use cases:

  • Business case one-pager: Enter the numbers a buyer shared in discovery and get a business case built on your approved ROI model, ready for a champion to forward to finance.
  • Mutual action plan: Give the buyer’s go-live date and get a backward-planned timeline with an owner from each side on every step.

Service Team Use Cases

Support teams sometimes first hear about a product change from a confused customer. A release-briefing skill converts engineering’s release notes into an internal briefing so agents are ready before launch.

What goes in the skill:

  • Instructions: Restate each change in customer-facing language, anticipate likely questions, and find every help article or saved reply that mentions the affected feature.
  • Reference files: Your terminology list and an exported index of help articles and saved replies
  • Template: What changed, potential customer questions, suggested replies, and content to update

Example input: “Release 4.12: CSV export moves from Settings to Reports. Exports over 50,000 rows now arrive by email.”

Example output:

  • Likely question: “Where did the export button go?”
  • Suggested reply: “Export has moved to the Reports menu. For files over 50,000 rows, we’ll email you a download link.”
  • Content to update: Three help articles and two saved replies still point to Settings.
  • Status: Draft, needs support lead review

Two more service use cases:

  • Bug report for engineering: Paste a customer’s problem description and get a ticket in engineering’s template, with reproduction steps and a severity level from your rubric. Clean handoffs prevent team collaboration from stalling on follow-up questions.
  • Quarterly business review deck: Add a customer’s usage export and support summary to get a draft QBR deck in your slide template, with talking points on adoption and risks.

Claude Cowork Best Practices for Safety and Reliability

Use caution with Claude Cowork because it can read, write, and delete files from the folders you point it to (with your permission, of course). I think this Claude Academy lesson paints the picture best with this analogy:

“Pointing Claude at Documents, Downloads, or Desktop is the equivalent of letting a new colleague rummage through every file you have. Make a folder for the work. Move (or copy) in what’s needed. Point Claude there.”

On top of the access you grant Claude through Cowork, skills downloaded from the internet can contain malicious code, so it’s important to know where the skill came from and review its content before installing.

This section covers the basics of how to keep your data safe while using Claude Cowork skills. For a deeper dive, I recommend reading Claude Support’s article on using Cowork safely.

Screen and Data Hygiene Before Recording

Everything visible during a recording is evidence for the model. Claude’s safety regulations recommend:

  • Close notification banners; sign out of unrelated accounts.
  • Hide API keys, tokens, and passwords.
  • Swap real customer data for a sanitized sample file before you start recording.
  • Close any files, apps, or conversations you don’t want recorded.

A skill built on a real spreadsheet can carry names, account numbers, or internal URLs into a document you later share with others. So, build the skill on filler data. Test it on real data only after you review it.

QA Checklist and Escalation Rules

Before you use, share, or deploy a skill to your team, check these items:

  • Triggers: Test several prompts that should invoke the skill, then review Claude’s thinking to confirm it loaded. You can also directly ask Claude in the chat, “Did you use the [skill name] skill?”
  • Security scan: On Enterprise plans with scanning enabled, skill and plugin scanning checks third-party skills and plugins for malicious content whenever they’re uploaded or edited. Skills already installed before scanning was enabled aren’t checked. Here’s what a security scan looks like on a recent skill I uploaded:
  • Approval mode: Cowork can be set to automatically approve or manually approve. Use “Manually approve” for sensitive accounts, a tool or plugin you haven’t tried before, or actions you can’t easily reverse.
  • Stop rules: If Claude opens files or sites you never mentioned, or the job outgrows your request, stop the task. When you build a skill, write stop conditions into its instructions as well. “If X is unclear, stop and ask before continuing” is a valid instruction.
  • Ownership: On Team and Enterprise plans, owners can require review before a skill is published. Assign one person to maintain each skill, since skills drift if nobody owns them. That person should verify it after product updates and version it before changes.

Connect Claude to HubSpot

Skills are useful for detailing how to do a task, but they can’t on their own get up-to-date information about your business and customers. So if a workflow depends on a deal stage or sales call notes, you would have to manually enter that yourself — unless you have a connector.

This is what the HubSpot connector for Claude does. It’s an official MCP (Model Context Protocol) connector built by HubSpot that gives Claude read and write access to your CRM — directly from the Claude chat window.

What the HubSpot Connector Does

The connector is a permission-based integration. You connect it once via OAuth; a Super Admin or user with HubSpot Marketplace permissions must authorize it in your HubSpot account first.

Once connected, Claude can:

  • Search contacts, companies, deals, and tickets in plain language
  • Create and update CRM records, including contacts and deals (currently capped at 10 records per bulk operation in beta)
  • Log activities, tasks, and notes directly from Claude
  • Surface insights from your engagement history — calls, emails, meetings

It also does not support delete operations. It also does not access HubSpot Sensitive Data properties. If your account has Sensitive Data enabled, it affects engagement data access as well — verify with your admin before building workflows on top of it.

Pro tip: Set the connector’s write tools to “Needs Approval” in Claude, as HubSpot recommends, so nothing changes until you confirm it.

Also worth knowing: according to HubSpot’s connector page, Anthropic does not use information shared through HubSpot to train its models, except when you choose to provide feedback or opt in.

Example CRM-Grounded Skills to Try First

1. Pre-call brief skill: Before a meeting, ask: “Pull everything we know about [Company] and give me a one-page brief.” The connector fetches the contact record, recent notes, open deals, and engagement history. The skill structures the output: background, open items, suggested talking points.

2. Post-call note and update skill: After a call, describe what happened in plain language. The skill formats a note in your standard structure. The connector logs it to the right deal or contact in HubSpot. No tab-switching required.

3. Pipeline review skill: “Which deals in Negotiation haven’t been touched in 10+ days?” The connector pulls the data; the skill formats it as a prioritized list with suggested next actions per deal.

4. Lead enrichment skill: When a new contact comes in, trigger a brief: company overview, LinkedIn context (if you’ve added a web search tool), recent news. Attach a suggested outreach opener. Log the research as a note.

These workflows live in the HubSpot connector for Claude — explore it here.

Frequently Asked Questions About Claude Cowork Skills

Do skills work in Cowork, Claude Chat, and Claude Code?

Yes, skills enabled in Cowork also work in Claude Code because they sync whenever you sign in with the same Claude account (as long as you have v2.1.273 or later). However, it’s a one-way sync — so skills you install in Claude Code do not sync with Claude Cowork or chat. Skills also work in Claude Chat. They’re a Claude-level feature, and the only Cowork-exclusive feature was Record a Skill, which still has some limitations since the September 2026 merger of Cowork and Chat.

Can I safely use Record a skill with customer data?

Use fake test data instead of real customer data when recording your skill. Anthropic warns against displaying sensitive information while recording, because anything visible on your screen or spoken aloud during the session is captured. Test with real data only after you’ve reviewed the generated skill carefully, since it can carry visible information into a reusable document.

How do I share skills with my team and control updates?

On Team and Enterprise plans, once an owner turns on sharing, open Customize > Skills, click the three-dot menu beside a skill you built, and choose Share. You can then add people by name or email. Recipients get a view-only copy that picks up your edits the next time they use it. Admins can control what’s published to the org directory, and recorded skills default to private until the creator shares them. Cowork keeps internal version history per skill, so you can roll back if an update breaks something. On Team plans, there’s a soft cap of 50 custom skills per workspace; Enterprise caps at 200.

All it takes is one skill to start.

← All articles

What this article says

More in MarTech & AdTech

All →
Social media style guide: Template, examples, and tips for 2026
Hootsuite

Social media style guide: Template, examples, and tips for 2026

Social media governance: 9 risks and how to avoid them
Sprout Social

Social media governance: 9 risks and how to avoid them

The Best SMS Marketing Platforms for 2027, Compared by Use Case- Klaviyo
Klaviyo

The Best SMS Marketing Platforms for 2027, Compared by Use Case- Klaviyo

Lead Magnet Ideas and Examples (70+ Ways)
Activecampaign

Lead Magnet Ideas and Examples (70+ Ways)

Welcome Email Series: A 6-Email Sequence
Activecampaign

Welcome Email Series: A 6-Email Sequence

8 Best Social Listening Tools in 2026
Sprinklr

8 Best Social Listening Tools in 2026

More from Hubspot

How to track your brand's visibility in AI search [2026]
Hubspot

How to track your brand's visibility in AI search [2026]

Claude AI for marketing: How teams use it with HubSpot
Hubspot

Claude AI for marketing: How teams use it with HubSpot

AI influencers: What consumer sentiment data tells brands
Hubspot

AI influencers: What consumer sentiment data tells brands

AEO benchmarking: What market data reveals about AI search
Hubspot

AEO benchmarking: What market data reveals about AI search

How to use AI to write a blog post (that isn’t slop)
Hubspot

How to use AI to write a blog post (that isn’t slop)

Branded vs. unbranded queries in AI search — and how to track them
Hubspot

Branded vs. unbranded queries in AI search — and how to track them