LLM usage

To build the initial version of this project, I experimented with an LLM model: Gemini Pro 2.5 (experimental). This page contains a description of how the experiment went, and what things I learnt.

How it went

• At first, I wasn't sure of what my needs were or what I needed to build. So I started by explaining Gemini what my initial needs were (building a Cadiretis clone for nostalgic reasons, and the fact that I needed to track my working locations), and asked it to help me create the project plan you can read at //docs/project_plan.
• Iteratively, we created the project plan section by section, sometimes reviewing previous sections as well. We finally reordered some of the information and sections in order to make it clearer for a first-time reader, since some concepts appeared before they were explained.
• Once that was done, I asked Gemini to generate the code directly. This resulted in a code soup, which I had to heavily refactor afterwards. Although the initial commit incorporates this refactor, some bad code was still left not refactored due to lack of time.

Things I've learnt

• Instead of asking the model to directly output code, I will try to instruct the model to follow a development cycle. For instance, I might want to incorporate TDD into the process.
• Next time I will make sure to ask for unit tests in the technical requirements.
• I asked the LLM to use Bazel as the build system, but I wasn't too much familiarized with it, so the road was bumpy. I learned it's better to learn about a tool before asking a LLM to use it (note for the future: maybe LLMs can still help in the learning process?), since the decisions that the LLM takes are not sometimes the best ones, and some issues I had would have easily been fixed if I had read the documentation beforehand.
• I quickly took over after the initial version, due to some issues with the code and the fact that it was painful to instruct the AI to fix the issues. I might want to try having a more educational approach with the model in the future.
• Related to the previous point: using Gemini's web UI for building a codebase was painful. I tried to use the canvas feature, but it was hard to sync local changes to the generated Canvas (so sometimes I kept the changes locally), and it took a lot of time for the model to regenerate it. Towards the end, Gemini started to create a lot of canvases so I ended up with duplicate files and a whole mess. Maybe AI agents (see avante.nvim) can help alleviate this pain point? It's something to research for the future.