Using AI as a framework

Lessons learned 1/11 after 15 years in software and 1 year of intensive Cursor use

Currently, AI — rather than being intelligent — behaves like a big framework. You can use its ready-made functions without worrying too much about what's inside. It will also use other frameworks to simplify its work and keep its code size down.

In short: you have a need, the AI patches together known solutions to solve it. And like any framework, you'd rather not need to dig into its internals.

The limits come quickly

Problems arise fairly quickly because the underlying technology (LLMs) is limited.

To understand how it works, imagine that the bigger a model is, the more constraints it can satisfy simultaneously. Like a game of pick-up sticks, the software engineer must respect a growing number of constraints that ALL need to be satisfied at the same time for the project to work.

This is why you can hit a ceiling fairly quickly when using AI on complex projects.

How to push the ceiling further

You can push back the problem in two ways:

1. Use only the relevant parts in the prompt context — isolate the necessary files
2. Use a bigger model that can handle more constraints simultaneously

Summary

• Use AI as a framework — ready-made functions without worrying about the details
• Use a large model if the project is complex
• Isolate the necessary files when possible to keep a clean context
• Switching from a large model to a smaller one → guaranteed disaster

Bonus: prompting in French

I write my prompts in French. First because I'm more precise, and because Claude performs just as well in French as in English. Switching to English improves response quality for some less robust models, but the biggest ones have no issues at all.

---

Originally published on LinkedIn.