Simon Willison's Weblog

This is the clearest working mental model for exactly your situation: treat the AI as a fast, over-confident pair programmer who looks things up instantly but will happily hallucinate a library that does not exist. Willison walks through how he actually prompts, plans, and takes over when the model drifts, so you learn to direct it instead of trusting it blindly. It is the single best answer to why you still have to read and test what comes out.

Here's how I use LLMs to help me write code

From Simon Willison's Weblog by Simon Willison 20 min read

• Think of the model as a very fast junior who never asks clarifying questions
• Always test the output yourself, LLMs hallucinate methods and libraries that look real
• The real win is shipping small things that were never worth your time before