[_tqr3]

I’ve tried building a web app with LLMs before. Two of them went in circles—I'd ask them to fix an infinite loop, they’d remove the code for a feature; I’d ask them to add the feature back, they’d bring back the infinite loop, and so on. The third one kept losing context—after just 2–3 messages, it would rebuild the whole thing differently.

They’ll probably get better, but for now I can safely say I’ve spent more time building and tweaking prompts than getting helpful results.

[diggan]

Rather than doing that approach which eventually builds up to 10+ messages or more, iterate on your initial prompt and you'll see better results. So if the first prompt correctly fixed the infinite loop, but removed something else, instead of saying "Add that back again", change the initial prompt to include "Don't remove anything else than what's explicitly mentioned" or similar, and you'll either get exactly what you want, or some other issue. Then rinse and repeat until completed.

Eventually you'll build up a somewhat reusable template you can use as a system prompt to guide it exactly how you want.

Basically, you get what you ask for, nothing else and nothing more. If you're unclear, it'll produce unclear outputs, if you didn't mention something, it'll do whatever with that. You have to be really, really explicit about everything.