Code will be generated, not written. Working with 100+ companies, I have seen this firsthand: most enterprise workloads taking dependencies on AI SDKs are usually net-new microservices. Hardly anyone has Windows-scale codebases for their new AI features. This is modular, greenfield work. Perfect for coding agents.

The catch? Out-of-the-box agents lack domain knowledge about your SDKs and patterns. But frontier LLMs are extraordinarily sample efficient. The patterns you need are already encoded in their latent space from pretraining, especially when working with popular languages. All you need to provide is the right activation context. That is what skills do.

What’s in Agent Skills?

github.com/microsoft/agent-skills is a collection of skills, prompts, agents, and MCP configurations that provide the activation context to specialize coding agents for Microsoft Foundry and Azure development.

The repo contains:

• Skills: Modular knowledge packages for specific domains (Cosmos DB, Foundry IQ, AZD Deployment, etc.)
• Prompts: Reusable templates for code reviews, component creation, endpoints
• Agents: Persona definitions (backend, frontend, planner)
• MCP Servers: Pre-configured servers for GitHub, Playwright, Microsoft Docs, Context7
• Context7 Integration: Indexed Foundry documentation at context7.com/microsoft/agent-skills

Systems Thinking First, Then Skills

Before loading skills into your agent, you need a mental model of what you are building, reducing your unknown-unknowns. Skills amplify your direction, they do not replace it.

My recommendation: do the basic hello world for each SDK you plan to use. For example, for Foundry IQ, walk through the agentic retrieval pipeline quickstart. Understand the abstractions at a high level. Know what Cosmos DB containers are, how Azure AI Search indexes work, what a Foundry agent does. You need this picture to prompt and steer the agent effectively.

Beyond that high-level understanding? You probably will not be hand-coding the idiomatic patterns for how you consume these services. The model already has these patterns encoded. Skills surface them. You bring the architecture and intent. The agent brings the implementation details that are already latent in its weights.

Avoiding Context Rot

Do not load all skills at once. Context rot occurs when your agent’s context window gets cluttered with irrelevant information, degrading output quality.

Loading everything will:

• Dilute the agent’s attention across unrelated domains
• Waste precious context window tokens
• Cause the agent to conflate patterns from different frameworks

Only copy the specific skills essential for your current project. High signal-to-noise ratio is everything.

The same principle applies to MCP servers. Do not enable every server in your config. Too many tools create the same dilution problem. Pick the servers relevant to your current task.

Quick Start

1. Copy Only the Skills You Need

Browse available skills and copy relevant ones to your project:

# Clone the repo
git clone https://github.com/microsoft/agent-skills.git

# Copy specific skills to your project
cp -r agent-skills/.github/skills/cosmos-db-python-skill your-project/.github/skills/
cp -r agent-skills/.github/skills/foundry-iq-python your-project/.github/skills/

Or use symlinks. For multi-project setups:

# macOS/Linux
ln -s /path/to/agent-skills/.github/skills/mcp-builder /path/to/your-project/.github/skills/mcp-builder

# Windows (PowerShell as Admin)
New-Item -ItemType SymbolicLink -Path ".github\skills\mcp-builder" -Target "C:\agent-skills\.github\skills\mcp-builder"

For sharing skills across different coding agents in the same repo:

# Symlink to multiple agent config directories
ln -s ../.github/skills .opencode/skills
ln -s ../.github/skills .claude/skills

2. Configure Your Agent

GitHub Copilot: Skills are auto-discovered from .github/skills/

Claude Code: Reference SKILL.md files in your project instructions or CLAUDE.md

GitHub Copilot CLI: Works with the same skill structure

3. Add MCP Servers

Copy the MCP configuration to your project:

cp agent-skills/.vscode/mcp.json your-project/.vscode/

Then remove the servers you do not need. This gives your agent access to:

• Microsoft Learn MCP: Ground agents in official docs
• Context7: Indexed Foundry documentation
• GitHub, Playwright, Terraform servers
• Sequential Thinking, Memory tools

4. (Optional) Use Context7 for Foundry Docs

For real-time Foundry documentation access, add the Context7 MCP server or use directly:

Context7 URL: https://context7.com/microsoft/agent-skills

A GitHub workflow runs weekly to sync the latest Foundry documentation updates automatically.

Available Skills

Scaling Beyond Static Skills

For larger codebases where static skills are not enough:

• AST-based Memory Systems: Hierarchical graph representation of code structure
• Context Graph: Knowledge graphs tracking component relationships and dependencies

These help agents maintain understanding without cramming everything into the context window.

What’s Next

This is v1. I will be adding more idiomatic skills for Microsoft SDKs, covering the cloud services and frameworks teams actually use.

I am also testing agent skills with copilot-sdk to validate skill quality programmatically. You can see a sample test harness at copilot-sdk-samples/skill-testing.

For now, start with github.com/microsoft/agent-skills, pick only what you need, and keep that signal-to-noise ratio high.

Resources

• Agent Skills Repository
• Context7: Foundry Docs
• Copilot SDK
• Claude Code + Foundry Setup Guide
• Codex + Azure OpenAI Integration
• Microsoft Foundry Docs