Microsoft's dotnet/skills library gives coding agents .NET expertise

A .NET developer runs a coding agent and watches it stumble on a project's build file. The agent writes code that compiles, then breaks the test runner, then fixes the test runner wrong. Microsoft posted a repository that aims to stop this specific loop.

The repository is dotnet/skills on GitHub. It ranked tenth on Trendshift's daily list today. The license is MIT. The star count is 6,041, read today from the GitHub API. The repository was created in February 2026 and the latest push was today. A nightly release of its skill validator went out the same day.

What the library covers

The repository bundles 16 plugins grouped by task. The dotnet plugin holds the main .NET skills with a C# language server. The dotnet-msbuild plugin covers build failures, build performance, and code quality. The dotnet-test plugin runs tests, measures coverage and migrates test frameworks. The dotnet-upgrade plugin migrates a project across .NET framework versions. The dotnet-ai plugin helps a team pick between an LLM, RAG, MCP or classic ML with ML.NET. There are also plugins for ASP.NET Core, Blazor, WinForms, NuGet, MAUI and Entity Framework.

Every plugin sits in its own folder under plugins/. Each one is a directory of skill files. Each skill has a SKILL.md that tells the agent when and how to use the skill. The agent host provides the credentials and the tools. The skill does not grant access to anything. A separate agentic-workflows/ folder holds reusable GitHub agent workflows, installable with the gh aw CLI.

How it fits with a working team

The skill format follows the Agent Plugins 1.0 specification from agent-plugins.org. That is the portable interface Google, OpenAI and Anthropic have agreed on. In practice, this means a .NET team can point Claude Code, GitHub Copilot or Codex at the same skill directory. The repository also carries a Skill Value dashboard at dotnet.github.io/skills. It shows token use, elapsed time, activation and not-passed rates by plugin, skill, executor model and judge model. That is the first public dashboard of its kind for a vendor skill set, that we have seen.

What this does not cover

The README is explicit about two gaps. The repository carries custom agent files under agents/*.agent.md. Those files follow GitHub Copilot conventions, not native OpenAI Codex conventions. Codex's own install path does not pick up these files yet. A Codex user has to wire them up by hand. The validator releases are nightly, which is Microsoft's own warning that the tool is still changing. No published benchmark attaches real numbers to how much time these skills save. The skills are new. The user counts are not disclosed.

What to do with this today

A .NET team using a coding agent can install a plugin in an afternoon. Pick one task that keeps breaking. The obvious places to start are test upgrades, MAUI builds and MSBuild diagnosis. Each one has its own plugin and a narrow SKILL.md. The cost of the test is a branch and a few hours. The upside is a documented skill set from Microsoft itself. The trade is a project that is still young and changes fast. Suppose your .NET team keeps hand-writing agent instructions for the same tasks. This is the moment to try the shared ones.

Source

dotnet/skills on GitHub.