Working with AI

I should perhaps say something about how Generative AI is used in the FrankenScore project. First of all, I have a prompt of about 4100 lines which prefaces every conversation with the AI chat client. The prompt consists of project documentation, background, design principles and goals, coding principles and conventions, explanations of central code and code examples. It also includes a major part of the source.

This allows the AI to:

1. Analyse and describe the project,
2. Converse about architectural choices,
3. Assist with project planning and open-source strategies,​
4. Judge how well the project embodies best Clojure practices,
5. Generate, with guidance, Clojure code and tests, and
6. Produce meaningful documentation and READMEs.

The copy on this website was almost entirely created by AI means, often using multiple iterations until I arrived at something suitable for publication. There remain a few passages that slipped me by as the AI produced text that reads a little too self-congratulatory on my part, but it was simply the opinion of the AI (though it is of course nice that it likes the code). I'll fix that during the days to come. Also, the technical comparison with other software is a bit too speculative and monotone. I'll change that, too.

In terms of code, I've found that Claude 3.5 Sonnet reasons better at depth about Clojure code than GPT-4o and consequently is the superior choice for complex coding. GPT-4o is still useful for producing text, though. It isn't exactly bad at coding, but it has a tendency to vomit code at you at every opportunity, which is both tiresome and expensive. Also, it kind of loses track when conversations get very long.

And they do; the chains of thought are sometimes complex, and a meandering AI can get costly. Therefore using Claude saves money in the long run.

By the way, it's easy to tell when I am writing. Just look for signs of British English. You know, -ise and colour and whilst and so forth. The AI invariably produces American English.

---