AI agent progress; refining, coding, training.

Spent a lot of time yesterday feeling pretty defeated and thinking about what to do next. Thought about trying the whole AI agent thing again. So I tried refining the prompt for the course and I seem to be getting okay results. Not perfect, but much better than before. Ironically, a locally running deepseek-r1:14b gives me better outlines (and actually returns proper yaml structure I told it to return) than both Claude Sonnet 3.7 and gpt4.1.

Maybe I could really try and give the whole AI agent thing a go again?

Did some more PROOMPTING and got to a point where my outline and chapter content prompts are returning something sensible. I'm now trying to set up the python cli but struggling.

I've set it up and got it to work how I want. It's actually looking pretty promising. I can generate an outline, generate content per chapter, and split it into the format I need for my importer. Next up, I need it to generate code for each step. This is probably going to be tricky, especially the part with ensuring each step has a logical progression from the last.

The code generation now works. Not sure how well it is, but it's something. I will probably need to get some training data to fine tune the model.

I created 15 examples of a nice outline using Cursor and converted it to training data for OpenAI. It's taking ages for them to finish training. Can't imagine how long it would take to train on 100+ examples. It's been 2 hours and it's still stuck on "Waiting...".

The tutorial content is also pretty shit, although in a correct format. It will definitely take a lot of fine tuning, but hopefully I can get it to a point where I only need to do minimal changes.