.NET Skills review Reproducing two build-diagnostic workflows
A pinned MSBuild skills review using shared output paths, corrected offline compilation, four independent binlogs and text replay, with clear limits on host and model claims.
AgentIndex · Published 2026-11-04 · Updated 2026-11-04
.NET Agent Skills organizes development tasks into concrete investigation steps. For an agent, knowing which evidence to collect first and when a technique is inappropriate can be more useful than another general instruction to improve code. The repository supplies skills and some host-specific components; the development environment still supplies the actual toolchain.
We selected two independently checkable MSBuild workflows: generating binary logs and investigating output-directory clashes. Both could be followed manually in our environment and produced inspectable evidence. We did not test automatic skill activation or measure model repair success, token savings or development speed.
How we tested the instructions
On November 4, 2026, we used commit 9837da685079d576feb27f9dd4f87631d3d94107, macOS arm64 and .NET SDK 10.0.102. Our fixture contains two handwritten minimal C# class libraries targeting net10.0, without additional NuGet packages. Pinned MSBuild skills
We first applied shared properties that sent both projects to common output and intermediate directories, then queried the evaluated MSBuild properties. Removing those overrides restored SDK defaults. We restored and compiled each corrected project, generated automatically named logs, and exercised the documented text-replay fallback. The reproduction fixture includes our verification script, source fingerprints and results.
| Check | Observed result |
|---|---|
| Pinned skill-file fingerprints | Matched recorded SHA-256 values |
| Deliberately shared output and intermediate directories | Evaluated paths matched across projects |
| Restore SDK defaults | Absolute output and intermediate locations became distinct |
| Corrected class libraries | Both restored from an empty local package source and compiled |
| Two restores and two builds | Four distinct, nonempty binlogs |
| Text replay of a build log | Project and OutputPath evidence available |
We did not force concurrent file corruption or claim to reproduce every intermittent build failure. The narrower questions were whether this configuration directs two projects to the same locations, whether the corrected projects still compile, and whether evidence can be retained and inspected.
Evaluated properties are the evidence
A string in a project file is not necessarily the final path. Shared props, configuration, target framework and global properties may contribute to evaluation. The skill compares actual paths and project instances rather than treating a search for one XML property as a complete diagnosis. Clash investigation skill
Our fixture also illustrates why adding a framework suffix cannot distinguish two different projects using the same framework. The full directory must preserve the necessary project boundary. Intermediate paths matter because restore artifacts are written there too. Separating final DLLs while retaining one shared intermediate directory is not the same correction.
In a real investigation, record project identity, configuration and relevant global properties. Normalize separators, resolve paths to absolute locations and compare actual destinations. An evaluation is not always a write: the skill explains exclusions for reference-resolution instances that only query metadata. We did not construct those advanced cases in this fixture, so their behavior is outside our execution result.
Log naming is part of reproducibility
The binary-log instructions use /bl:{} separately for each MSBuild invocation, allowing supported versions to choose unique names. Our two restores and two builds generated four independent files. The placeholder requires a sufficiently recent MSBuild or .NET 8 SDK and later; our test used .NET 10. Binlog generation skill
Retaining logs helps an agent explain a change using the original evidence rather than infer the earlier state from the final successful build. A log's existence is only the first check, however. The investigator still needs to identify the command, configuration and participating project represented by that file.
Our fixture checked that the files existed and were nonempty, then asked MSBuild to replay a selected binary log into diagnostic text. It did not interpret the binary itself as an ordinary text file. This fallback worked with the tested SDK; we did not establish compatibility with every historical SDK or operating system.
Separate skills, hosts and tools
The repository distinguishes portable skills from host-specific components. Agents, language services and MCP servers declared in a plugin require their own host support. One working component does not demonstrate that the complete package has been integrated. Component overview
For example, the pinned MSBuild plugin declares a binlog MCP launched through dotnet dnx with a prerelease option. We did not install or invoke that server; we tested the documented text fallback instead. Teams choosing the MCP path should verify the resolved package version, host connection and available tools independently. Plugin manifest
Start adoption with one task the team actually repeats. Confirm its SDK and workload requirements, then execute the described steps in a controlled project. Next, test whether the host discovers and activates the skill and follows the complete process. This separates a missing environment dependency from an activation failure or an incorrect diagnostic conclusion.
That distinction also makes evaluation more useful. A model might follow the instructions correctly but lack the required tool, or collect the right evidence but draw the wrong conclusion. Record the intermediate observations so those outcomes can be distinguished rather than reducing them to one undifferentiated success flag.
Adoption decision
The repository is a worthwhile candidate for developers maintaining .NET applications who want repeatable investigation practices. Begin with a relevant plugin and a small task. It does not replace SDKs, platform workloads or business tests. We did not evaluate MAUI, Windows UI, framework upgrades or the entire collection.
The pinned repository uses MIT licensing. A useful acceptance artifact is a repeatable fixture, corresponding logs and a conclusion tied to evidence. The number of installed skills is a much weaker indicator of value. License
For application preparation, see NocoBase 3. For a different example of preserving evidence and unknown results, read our Claude Ads review.