// the find
microsoft/skills
Skills, MCP servers, Custom Agents, Agents.md for SDKs to ground Coding Agents
A Microsoft-maintained collection of 174 'skills' — structured markdown files that prime AI coding agents with domain knowledge about Azure SDKs, Foundry, and related services. Think of each skill as a SKILL.md that tells Copilot or Claude exactly which APIs to use, which patterns to follow, and which anti-patterns to avoid, so you don't get hallucinated method names. Aimed at teams building on Azure who are already using AI coding agents.
The test harness is the standout feature — 1158 test scenarios that actually run the Copilot SDK against generated code and check acceptance criteria, not just 'does it compile'. The plugin model (npx skills add) with symlinks for multi-agent setups is practical: one skills directory shared across Copilot, Claude, and other agents without duplication. Coverage across five languages (Python, .NET, TypeScript, Java, Rust) with consistent naming conventions means the pattern is transferable across stacks. The daily GitHub Actions workflow that scrapes docs and regenerates llms.txt files means the skills stay current without manual updates.
Heavy Azure lock-in — this is not a general agent skills framework, it's Microsoft's skills for Microsoft's services; if you're not already in the Azure ecosystem, almost nothing here applies. The 'use skills selectively' warning in the README is real: loading the wrong set costs you tokens and can actively hurt generation quality by conflating patterns, but there's no built-in mechanism to enforce selective loading beyond developer discipline. The Foundry skills (hosted agents, toolboxes, IQ knowledge bases) are all marked preview, so a non-trivial chunk of the catalog may break or be deprecated under you. The symlink-heavy structure (skills/ at root symlinking to .github/plugins/) adds friction on Windows where symlinks require elevated permissions or developer mode.