Anthropic Launches Skills and Microsoft 365 Integration for Claude: Transform Your AI Assistant Into a Specialized Expert

Anthropic just dropped two major updates: Agent Skills and Microsoft 365 integration. Here's what you need to know.

Claude Agent Skills: Give Claude superpowers on demand

Skills are folders with instructions that Claude loads only when relevant to your task. Think of them like onboarding guides—you're teaching Claude your workflows, brand guidelines, or specialized expertise without cluttering its memory.

The concept is brilliantly simple: instead of building separate custom agents for every use case, you create reusable folders that Claude discovers and loads dynamically. Alex Albert says it's like Neo learning Kung Fu in The Matrix... Claude gains specialized knowledge instantly, but only when it's actually needed.

How they work:

• Progressive disclosure: Claude first scans skill names and descriptions at startup, then loads full instructions only if the skill is relevant to your task, then accesses additional bundled files as needed.
• Composable: Skills can stack together for complex workflows (e.g., using Excel + presentation skills simultaneously).
• Portable: They work everywhere—Claude.ai, Claude Code, and the API—using the same format.
• Powerful: Skills can include executable code for tasks where traditional programming is more reliable than token generation (like sorting massive lists or extracting PDF form fields).

You've actually already seen Skills in action if you've used Claude to create spreadsheets, presentations, or PDFs. Those capabilities? They're powered by Skills that Anthropic built. Now you can create your own.

Getting started with Skills:

1. Go to Settings > Capabilities > Skills in Claude.ai (available for Pro, Max, Team, and Enterprise users).
2. Enable the "skill-creator" skill, which provides interactive guidance. Claude asks about your workflow, generates the folder structure, formats the SKILL.md file, and bundles resources for you.
3. Or browse example skills on GitHub that you can customize for your needs.

What you can build:

• Company-specific brand guidelines and tone of voice
• Custom data analysis workflows
• Industry-specific templates (legal document formatting, financial modeling, etc.)
• Technical documentation standards
• Sales scripts and proposal frameworks

The technical structure is straightforward: a skill is just a directory containing a SKILL.md file that starts with YAML metadata (name and description), followed by instructions. For more complex skills, you can bundle additional markdown files, scripts, images, or any other resources Claude might need.

Skills can include executable code, so Claude can run scripts directly instead of burning tokens trying to generate solutions. This makes workflows both faster and more reliable. Just stick to trusted sources when installing skills, since they do give Claude code execution capabilities.

For developers, Skills are now available through the Messages API and a new /v1/skills endpoint. Claude Code users can install skills via plugins from the anthropics/skills marketplace, or manually add them to ~/.claude/skills.

Microsoft 365 + Enterprise Search: Your company knowledge, unified

For Team and Enterprise customers, Claude now connects to Microsoft 365 through a new MCP connector. This is a game-changer for organizations that live in the Microsoft ecosystem.

What Claude can access:

• SharePoint and OneDrive: Search and analyze documents across sites and libraries—from project specifications to strategic plans—without manually uploading anything to Claude
• Outlook: Access email threads, analyze communication patterns, and extract insights from your correspondence to understand project status, client feedback, or team alignment
• Teams: Search through chat conversations, channel discussions, and meeting summaries to surface decisions, track project updates, and understand what your team has been discussing

Even better: Enterprise search creates a shared, dedicated project for your entire organization that pulls from all connected tools. This project is personalized with your company name and includes custom prompts to help Claude search effectively across your company's apps.

When you ask Claude a question, it searches across your company's connected data sources in one place. Ask about your company's remote work policy, and Claude synthesizes information from HR documents in SharePoint, email discussions in Outlook, and team guidelines from various sources into one comprehensive report.

Real-world use cases:

• Onboarding new employees: New hires can ask Claude about company processes, policies, and tribal knowledge without bothering their teammates
• Strategic analysis: Ask Claude to analyze patterns in customer feedback across emails, Teams chats, and stored documents
• Finding experts: Quickly identify who in your organization has expertise on specific topics based on their communications and documents
• Project status updates: Get comprehensive overviews without manually checking multiple platforms

The setup is admin-controlled. For the M365 connector, administrators must first enable it for the organization before individual users can authenticate with their accounts. For Enterprise search, admins customize the shared project and curate which data sources to include.

Why we like this: 

Skills transform Claude from a general assistant into whatever specialist you need, while M365 integration ensures it has access to your actual company context. Combined, you get a truly customized AI that knows both how to work (Skills) and what to work with (your organization's data).

It's like having an expert employee who's already completed onboarding, memorized your style guide, and has instant access to every company file and conversation. No more "let me find that document" or "I need to check three different platforms"—just ask Claude, and it pulls together everything you need.

This is the direction AI is heading: not just smarter chatbots, but context-aware assistants that understand your specific workflows and have seamless access to your organization's collective knowledge.

Check out Skills documentation and M365 integration details to get started.

Now, the question is... can you replicate the functionality of Agent Skills in ChatGPT? Let's try! 

How to Replicate Claude Skills in ChatGPT

Great news: ChatGPT has several built-in features that can replicate much of what Claude Skills does—Custom GPTs, Projects, and Custom Instructions. Here's how to use each one:

Option 1: Custom GPTs (Best for Specialized Tasks)

Custom GPTs are personalized versions of ChatGPT that you can create without any coding. This is the closest equivalent to Claude Skills for specialized, reusable capabilities.

How to create a Custom GPT:

1. Go to chatgpt.com/create and log in.
2. Click "Create" in the top-right corner and enter your instructions in the message bar.
3. Under the Configure tab, you can provide detailed instructions on how the GPT should behave, add a description, upload knowledge files, and enable capabilities like Web Search, Image Generation, and Code Interpreter.
4. Click Create and select how you want to share your custom GPT, then click Update.

What you can include:

• Instructions defining behavior and functionality.
• Knowledge files that provide additional context for the GPT to reference.
• Capabilities like web search, code execution, and image generation.

Full guide on Custom GPTs →

Key difference from Skills: Custom GPTs need to be manually selected each time you want to use them, whereas Claude Skills load automatically when relevant.

Option 2: Projects (Best for Ongoing Work)

Projects are smart workspaces that keep chats, files, and custom instructions in one place, with built-in memory so ChatGPT remembers context.

How to use Projects:

1. Click the + icon next to "Projects" in the top-right corner, name your project, and hit "Create Project."
2. Click the three dots on the upper right corner to add project-specific instructions.
3. Upload reference files that ChatGPT can draw from when responding.
4. Move existing chats into the project by dragging them onto it or using the chat's menu.

Project-specific instructions examples:

• "Act like my marketing mentor. Be concise. Use bullet points. Ask clarifying questions."
• "Focus on data cleaning and visualization tips using Python."
• "You're a web developer helping someone with no experience create their personal website."

Key advantage: Project instructions override your global custom instructions and apply only within that specific project.

Full guide on Projects →

Option 3: Global Custom Instructions (Best for Consistent Preferences)

Custom instructions allow you to share anything you'd like ChatGPT to consider in its responses, and these will be added to all new conversations going forward.

How to set up:

1. In Settings, select "Customize ChatGPT" or "Personalization" and click on "Custom Instructions."
2. You'll see two text boxes—one for preferences about yourself and one for how you want ChatGPT to respond.
3. The text fields have a 1500 character limit.

What to include:

• Your role or identity (e.g., "I'm a software developer that primarily codes in Java").
• Your preferences (e.g., "Write efficient, readable code that includes clear, concise comments").
• Context about your work, products, or common tasks.

Official guide on Custom Instructions →
Advanced templates and examples →

Combining Approaches for Maximum Effect

The best replication of Claude Skills uses all three features together:

1. Global Custom Instructions for your general preferences and role
2. Projects for specific ongoing work with dedicated context and files
3. Custom GPTs for specialized, reusable workflows you want to share or access quickly

Example workflow:

• Set global instructions for your writing style and preferred tone
• Create a "Marketing Campaign" project with brand guidelines, target audience docs, and campaign-specific instructions
• Build a Custom GPT for "Brand Voice Checker" that ensures content matches your style guide

Limitations Compared to Claude Skills

• No automatic loading: You must manually select Custom GPTs or enter Projects, whereas Claude Skills load automatically when relevant
• Less progressive disclosure: ChatGPT loads all project files and instructions at once, while Claude Skills only load what's needed
• No skill stacking: ChatGPT doesn't automatically combine multiple Custom GPTs the way Claude can stack multiple Skills for complex workflows

Bonus: OpenAI Agent Builder

OpenAI recently launched Agent Builder, a visual drag-and-drop node-based builder that supports MCP servers and allows you to build custom agents with multiple nodes like Agent, MCP, and Guardrail. This is a more advanced option for developers who want to create complex multi-step workflows.

So what if you could recreate Claude's Skills feature as an actual Agent Builder workflow?

The concept = we build a meta-agent that decides which specialized sub-agents (skills) to invoke based on the user's request.

How it would work:

Node 1: Start Node

• The start node acts as an entry point, offering input variables to represent user-provided text.
• User submits their request (e.g., "Create a sales report for Q4").

Node 2: Guardrails Node

• Guardrails provide an open-source, modular safety layer that helps protect agents against unintended or malicious behavior.
• Filters out inappropriate requests before processing.

Node 3: Skill Router Agent

• A decision-making agent that analyzes the request and determines which "skill" (specialized agent) is needed.
• Prompt (basic; definitely flesh this out): "Analyze this request and determine which specialized capability is needed: Excel analysis, presentation creation, PDF manipulation, brand guidelines, or general assistance."
• You can add a classifier "state changer" node after this prompt to set the state variables for each category that then become persistent across the entire workflow and accessible throughout.

After you create the skill router agent, you'll add an if / else node that creates branching paths (with if/else logic) to follow one of the four below paths, based on which "variable" the router selects for a given prompt. You might even want to add a while loop node if a user submits multiple tasks, so it properly routes all requested tasks per a given input to the appropriate "skill agent." 

Node 4-8: Specialized Skill Agents (branching paths; you'll want to set separate prompts for each agent for how best to tackle each task)

• Excel Agent: Connected to Code Interpreter, trained on spreadsheet best practices.
• Presentation Agent: Instructions for creating compelling slides, visual hierarchy.
• PDF Agent: PDF manipulation workflows and form-filling logic.
• Brand Agent: Loaded with your company's brand guidelines and tone of voice.
• Writing Agent: Specialized in content creation with your preferred style.

Node 9 (or add to each individual agent as tool calls): File Search/Knowledge Retrieval

• Built-in tools allow the model access to external or internal data with very low effort.
• Each specialized agent can access uploaded knowledge files (templates, examples, best practices).

Node 10: Output Formatter

• Consolidates results from whichever skill agent was triggered.
• Ensures consistent output format regardless of which path was taken.

Node 11: End Node

• Returns the final result to the user.

The advantage of this approach:

Unlike Custom GPTs that require manual selection, Agent Builder's visual canvas reduces development time and can cut iteration cycles by up to 70%. Your "Skills Agent" would:

1. Auto-route requests - Users don't need to know which skill they need.
2. Stack capabilities - The router could trigger multiple skill agents in sequence (e.g., Excel analysis → Presentation creation).
3. Version control - Agent Builder supports full versioning, ideal for fast iteration.
4. Easy updates - Modify individual skill nodes without rebuilding the entire system.
5. Export to code - The "Code" button generates implementation code that you can export to Python or TypeScript through the OpenAI Agents SDK.

Taking it further:

Progressive loading simulation: Create a "Skill Loader" node that:

• First checks a lightweight metadata file for each skill (name + description).
• Only loads the full instructions when a skill matches the request.
• Keeps the context window lean, similar to Claude's progressive disclosure.

MCP integration: You could also add external "skills" to your skill agents via the Connector Registry consolidates data sources like Dropbox, Google Drive, SharePoint, and third-party MCPs, so your skill agents could access company files and tools automatically.

Custom evaluation: Don't forget to set custom graders to determine whether the agent (both the main agent, and any individual skill agents) is performing to your expectations on your specific use case - test if the router is selecting the right skills and if outputs match quality standards. The more skills you add, the more complicated this will get, so evals are crucial. Watch this video to help set up your evals:

The result: A ChatGPT workflow that mimics Claude's Skills by automatically detecting what expertise is needed and loading it on-demand, without the user having to manually select anything. It's not as elegant as Claude's native implementation (you might want to bug OpenAI to add a similar feature!!), but with enough tweaking, you can get it remarkably close... and you can customize it endlessly.

As you can see, while ChatGPT doesn't have automatic skill-loading like Claude, combining Custom GPTs, Projects, and Custom Instructions (or even Agent Builder) gets you surprisingly close. You just need to be more intentional about organizing and selecting your specialized contexts.