ai in the flow
This week, I had a chance to just grind out some code and actually have AI get in the groove with me. It was an existing project that needed some backend and database work, and I was able to have AI help me with some of the work that was needed.
I had worked with the database before, but not extensively. To get started, I scripted out the database schema for myself (I was not sure exactly what I would be working on), then decided I would turn Gemini 2.5 Flash (I’ll refer to this as Flash from here on out) loose on it while I started working on some of the base queries/stored procedures. I chose Flash because I have always been impressed with its capabilities and speed. Gemini 2.5 Pro is good for more abstract tasks, but it can be very slow. For relatively simple tasks, Flash is more than capable and much faster. I didn’t want to wait around for an AI to just verify code validity and offer suggestions.
One of the first tasks I had Flash do was complete a Merge statement for me.
I had the basics of the query, but I always get bogged down in the syntax.
That was quick and easy for Flash to handle. Of course, I checked the query
to make sure it was doing the right things. I liked it, the IDE liked the syntax,
and we moved on.
I was actually saving all of the procedures I was writing down to a local directory (with the schema script) so it was easy for me to start Flash off in that directory and have it create a GEMINI.md file for me/itself. I didn’t have permissions in the database to create/execute stored procedures, so having Flash in there to periodically review and suggest changes was a big help.
If I would have known I wasn’t going to have database permissions, I would have started off using Visual Studio (with GPT-4o) instead of SSMS for the initial work. Predictive code completion is a huge time saver, and I would have loved that for some of the queries I was writing. But, once I had the basics of the procedures down, I was able to switch over to Visual Studio and have GPT-4o start to help get the repositories (controller/service/repository) set up.
I put most of the validation there before moving it to the service layer. GPT-4o does a pretty good job of generating the validation code, but it’s not a mind reader so I usually had to tweak it a bit. Throwing the right errors with the right messages it usually got right.
The models had already been created, and didn’t quite match up with the database schema, so I there was some back and forth with that. Flash did a good job of helping me with the procedures. I would remove some parameters I was expecting, and Flash would adjust the procedure accordingly or suggest changes or point out default values I had missed.
Once the repositories and the procedures were sync’d up, I was able to start on the service and controller layers. GPT-4o shined on the service layer. Setting up the repositories’ interfaces almost wrote themselves. It was a short and easy process of verifying code predictions from the implementation. Moving those interfaces to the service layer was just as straightforward as were those implementations.
The controllers were a bit more work, there wasn’t a really good pattern to set up. There were things that were just different enough that it was easier for me to write them myself than to try to get GPT-4o or Flash to do it. And really, it felt odd after having so much help with the rest of the code.
The last thing I wanted to do was move all of the validation from the repository to the service layer. I started by asking Flash to do it on a sample repository, and it did a pretty good job, but I didn’t want to refine the prompt and still triple check the code, so I just did it myself. That gave me a chance to check that the validation that was written was still the validation I wanted, and it was a good way to wrap up the project. GPT-4o was still able to help move and check the validation and make sure everything was in order.
Afterwards, I had Flash run over everything and it suggested a concurrency issue
in a procedure that I had missed. I suggested it change one of the statements in
there to a Merge statement (detect a trend), and it did so without any issues.
Overall, it was a very productive session with a smooth flow of AI and code.
No hiccups, no puzzles, just a steady stream of productivity. It wasn’t complicated
code, bits were tricky, but nothing incredibly difficult. AI was able to help
and stay out of the way. Sessions like this are incredibly satisfying.