I'm on my second cup of coffee and things are still a bit foggy. Microsoft Copilot isn't quite giving me the output I want or maybe I'm being over critical. I didn't feel like typing this morning, so I decided to chat with it to have it transcribe my thoughts. That didn't help. Hence, I am winging this post.
I'm constantly experimenting with different ways of increasing my productivity by using AI and I've been quite pleased thus far. There are some very clunky things like trying to copy and paste the markdown that Copilot generates into a LinkedIn post (doesn't work) or Substack (does work). Or generating some deep research from Google Gemini that I have to export to Google Docs and then copy and paste into Word to preserve the formatting (and citations).
I don't believe in just talking about the tools. I need to dive in deeply and actually use them. Not in a "Hello World!" sense but to actually create something useful. I've yet to try out the "Rewrite with AI" button on LinkedIn but be on the lookout for an alternate version of this article using the feature.
The latest experiments in our AI test lab have been with code generation tools like Gemini Code Assist, Cursor, and now Claude Code. Warp was supposed to be next but on a recommendation from a colleague, I decided to give Claude Code a try.
For this test, I went off script and didn't carry over the coding rules and prompt file. Also, after the last two experiments, I had discovered a few things about my original architecture which were off, so I made some adjustments. Regardless I gave Claude Code my prompt which included some of the original requirements that the other two ignored like "create unit tests for the front-end and back-end code" and "use .NET 8.0".
Claude Code fared well. It obeyed my instructions on creating unit tests and using SOLID, DRY, and YAGNI. It understood what use OOP meant and created interfaces. However, it fell down on creating a working app much less than one that actually builds. Yes, the React+TypeScript front-end and the Azure Functions back-end. Would. Not. Build.
While doing other things around the office, I spent time with Claude Code "vibe coding" (babysitting?). After I coaxed it into fixing the build issues, I had it go back for some refactoring which included adding XML comments, creating a Visual Studio solution file, removing underscores from fields, and finally dockerizing the front-end and back-end projects.
After another couple of hours of having Claude Code fix issues with the Docker builds, mock services, and npm package installs, I finally had a working application that allows me to access LM Studio which is running on my high-end gaming machine from a browser interface.
This morning, I'm back in the lab to start the next stage of coding which will be integrating the MCP servers into the fold.
Stay tuned for more...