Teaching Your AI Coworker

Monday AI Dispatch | June 1, 2026

Two of the most significant agentic AI platforms available right now, Claude Cowork from Anthropic and Copilot Cowork from Microsoft, both support custom skills and can be seen as an AI Coworker. That’s not a coincidence. It’s where enterprise AI is heading: from tools that respond to tools that learn your workflow.

Both platforms use a `SKILL.md` file as the core mechanism and both use YAML frontmatter for metadata. Both pick up skills automatically once deployed. And yet the experience of building a skill in each platform, what it can do, where it lives, and how reliably it fires, is meaningfully different. This post walks through both in detail; with a real example you can use today.

What a “Skill” Actually Is

Before comparing platforms, let’s agree on the definition.

A skill is a reusable, structured set of instructions that tells an AI agent how to perform a specific workflow consistently. This is without you retyping your preferences every time. It is not a prompt. A prompt is what you type. A skill is what gets loaded silently in the background to shape how the AI interprets and executes your prompt.

Think of it this way: you could tell a new employee how to format a meeting recap every single time they write one, or you could give them a style guide once. Skills are the style guide.

Both Copilot Cowork and Claude Cowork have converged on the same file format to describe skills. That’s the `SKILL.md` pattern: a Markdown file with a YAML header containing a `name` and `description`, followed by freeform instructions. This format originated from Claude Code’s skills system and has since been adopted by Microsoft for Copilot Cowork.

The Example Skill: LinkedIn Profile Reviewer

Rather than use a generic brand guidelines example, let’s build something more immediately useful: a skill that reviews a LinkedIn profile and provides structured feedback.

This is a skill almost anyone can relate to. Whether you’re a consultant trying to sharpen your positioning, a job seeker updating for a search, or a manager helping a team member build presence. A LinkedIn profile review is a repeatable workflow with clear criteria and consistent deliverable expectations.

Here’s the full `SKILL.md` for both platforms (the content is the same; what changes is how you deploy it):

---
name: LinkedIn Profile Reviewer
description: Review a LinkedIn profile and provide structured, practitioner-level feedback on headline, About section, experience narrative, skills, and professional positioning. Deliver output as an annotated critique with prioritized action items.
---
## Purpose
When asked to review a LinkedIn profile, analyze it against the following criteria and produce a structured written critique with a prioritized action list. Do not provide generic advice. Be specific about what is weak, why it is weak, and what a stronger version would look like.
## Review Framework
### 1. Headline
- Does it state value delivered, not just a job title?
- Is it optimized for the audience (recruiters, prospects, peers)?
- Does it use industry keywords without reading like a keyword dump?
### 2. About Section
- Does it open with a strong hook in the first two lines (visible before "see more")?
- Does it speak to outcomes and credibility, not just responsibilities?
- Does it end with a clear call to action or contact invitation?
- Is the writing voice consistent and genuine — not AI-generated-sounding?
### 3. Experience Section
- Do bullet points lead with action verbs and quantified outcomes?
- Is there a coherent narrative across roles, not just a list of jobs?
- Are recent roles weighted more heavily in depth?
### 4. Skills and Endorsements
- Are top skills aligned with the person's stated professional focus?
- Are outdated or irrelevant skills cluttering the list?
### 5. Overall Positioning
- Is there a clear, distinct professional identity?
- Would a stranger in this person's target audience immediately understand their value?
## Output Format
Produce your output in the following structure:
**Overall Assessment** (2–3 sentences summarizing the profile's current state)
**Section-by-Section Critique** (use the framework above; be specific and direct)
**Top 5 Priority Actions** (numbered list, highest-impact items first)
**One Example Rewrite** (take the weakest single element and show what it could look like)
## Tone
Direct, honest, and constructive. Do not soften critique with filler phrases. The person asked for a review, not validation.
## Example Input
"Here is my LinkedIn profile URL / text. Please review it using the LinkedIn Profile Reviewer skill."
## Example Output Structure
> **Overall Assessment:** Your headline buries the lead and your About section doesn't survive the two-line preview test. The experience section is strong on titles but weak on outcomes. Here's what to fix first.
>
> **Headline:** "Director of Technology | Microsoft Teams | SharePoint" tells me your job, not your value. Consider: "I help enterprise teams deploy Microsoft 365 Copilot that actually sticks — practical adoption, not just licenses."
>
> ...
>
> **Top 5 Priority Actions:**
> 1. Rewrite the headline to lead with the outcome you deliver, not your title.
> 2. Replace the first paragraph of About with a two-sentence hook that names your audience and your differentiation.
> ...

That’s the skill content. Now let’s talk about how to deploy it in each platform. Also, I will focus on where the experiences diverge.

Deploying in Claude Cowork

If you are looking for the easy path, scroll to Cowork section where you have Claude do the work for you!

Step 1: Create the directory structure

Claude Cowork discovers skills from a specific folder on your local machine (or in the Cowork file system). Create this structure:

~/claude-skills/linkedin-profile-reviewer/skill.md

The folder name should match or closely reflect your skill name. The file inside must be named skill.md (lowercase).

Step 2: Write your skill.md

Copy the content above into `skill.md`. The required YAML frontmatter fields are `name` and `description`. The description is especially important as Claude uses it to determine when to invoke the skill automatically. A vague description means the skill won’t fire reliably.

Keep the description under 200 characters and be precise: “Review a LinkedIn profile and provide structured practitioner-level feedback…” is better than “Help with LinkedIn.”

Step 3: Package and upload

1. ZIP the folder: `linkedin-profile-reviewer.zip` containing the `linkedin-profile-reviewer/` folder with `skill.md` inside.
2. In Claude, go to **Settings → Skills**.
3. Click **Upload skill** and select your ZIP file.
4. The skill appears in your skills list. Toggle it on.

Step 4: Test it

Open a new Claude conversation and type:

“Here’s my LinkedIn profile text. Please review it.”

You should not need to invoke the skill by name. If your description is accurate, Claude will recognize the intent and apply the skill automatically. If it doesn’t, try: “Use the LinkedIn Profile Reviewer skill to analyze this.”

In Claude Cowork specifically

In Cowork, there’s also a conversational path: run through a review workflow manually once, then at the end say:

“Package what we just did into a skill.” or “Create me a skill”

The built-in skill-creator captures your workflow, steps, and source locations so the next run starts from a single prompt. This is a notably low-friction entry point for people who don’t want to write a SKILL.md file from scratch.

Skills in Cowork are also available across surfaces the same skill that produces a structured critique in chat can produce a formatted Word document in Cowork, or a structured breakdown in Excel if the context calls for it. Claude adapts the output format to the surface, while the skill’s logic stays consistent.

Skills are available on Free, Pro, Max, Team, and Enterprise plans. Code execution must be enabled. For Team and Enterprise, admins can provision skills organization-wide.

Deploying in Copilot Cowork

Step 1: Create the OneDrive folder structure

This is the key difference from Claude: Copilot Cowork reads skills from OneDrive, not your local machine. Local files are not accessible. Navigate to your OneDrive and create:

Documents/
  Cowork/
    Skills/
      LinkedIn-Profile-Reviewer/
        SKILL.md

Create the Cowork folder if it doesn’t exist, then the `Skills` subfolder, then one subfolder per skill.

One important constraint: any templates or files your skill references must also live in OneDrive or SharePoint. Copilot Cowork has no access to local file paths.

Taking the Easy Way out

Just like Claude, you can also ask Copilot Cowork to create skill.

Step 3: Save and wait for discovery

There is no upload button or toggle. Copilot Cowork automatically discovers new skills from your OneDrive folder at the start of each new conversation. Save your `SKILL.md` file to OneDrive, start a new Cowork conversation, and the skill is available. You can also add up to 50 custom skills per user.

Step 4: Test it

Open Microsoft 365 Copilot, select **Cowork**, and type:

Critique my LinkedIn

Cowork will create a plan, display what it intends to do step by step, and ask for approval before any action that can’t be undone (like saving a document or sending a message). Unlike Claude, Cowork surfaces its reasoning process more explicitly and requires approval at action points.

Copilot Cowork requires a Microsoft 365 Copilot license and is currently in the Frontier preview program. Your tenant admin must enroll in Frontier via Copilot → Settings → Frontier in the M365 admin center.

Comparison

Where They Are Similar (and Why That Matters)

The SKILL.md as a shared format is not accidental. Both platforms are building on the same insight: skills are better than prompt libraries because they’re invoked automatically based on intent, not by users manually selecting them. The description field in the YAML frontmatter is essentially a classifier as the AI reads it to decide whether your current task matches the skill’s purpose.

This means writing a good description is the most important part of skill authoring in both platforms. A bad description produces a skill that never fires. Both platforms’ documentation emphasizes this, and both recommend being specific about the task domain rather than writing broad, aspirational descriptions.

Where They are Different

File location: Claude’s local-first approach means faster iteration; edit the file, zip, upload, test. Copilot’s OneDrive approach enforces collaboration naturally: the skills folder is already in a shared cloud environment, making team distribution straightforward but making individual iteration slightly slower.

Copilot Cowork’s explicit approval workflow is a deliberate design choice aligned with Microsoft’s enterprise governance posture. Before Cowork sends an email, posts to Teams, or modifies a document, it surfaces what it plans to do and waits for a green light. For workflows (HR communications, client-facing documents, financial data), this is a feature. Claude Cowork’s behavior is more contextual as it will act autonomously on lower-stakes tasks but surfaces checkpoints depending on the workflow.

The integration ecosystem. Copilot Cowork has deep, native M365 integration as it can draft in Word, schedule in Calendar, post in Teams, and pull from SharePoint with no configuration. Claude Cowork’s connector model is broader (Slack, Salesforce, Jira, Gmail, Google Drive), but relies on those connectors being set up. If your work lives in Microsoft 365, Copilot Cowork has an integration advantage by default. If your work spans multiple platforms, Claude’s connector breadth is more useful.

Summary

The skill-as-SKILL.md pattern is clean and standard. The friction to get a first skill deployed is low on both sides as it is under 15 minutes from intent to working skill in either platform.

The practical choice between them right now comes down to where your work already lives and what your license situation is. If you’re in a Microsoft 365 Copilot-licensed tenant enrolled in Frontier, Copilot Cowork’s native M365 integration is compelling and the custom skill system works well. If you’re on a Claude plan with Cowork access, the faster iteration loop, broader connector ecosystem, and executable code support in skills give it an edge for complex, cross-platform workflows.

Neither platform has finished building. Both are moving fast.

Related Links

• Cowork steps off the desktop – Pat Petersen
• Let Copilot Cowork Dig You Out – Pat Petersen
• Copilot Cowork – From conversation to action (Microsoft 365 Blog)
• Copilot Cowork – A new way of getting work done (Microsoft 365 Blog)
• Copilot Cowork FAQ (Microsoft Learn)
• How to Get Your Microsoft 365 Tenant Ready for Copilot Cowork (Governance-first guide)
• Claude Cowork product overview (Anthropic)

Have a skill you’ve built that others should know about? Drop a link in the comments.
 

The post Teaching Your AI Coworker appeared first on Pat Petersen.

Original Post https://patpetersen.com/2026/06/01/ai-coworker/