Large language model LLM

A large language model is trained to continue text. Everything else it appears to do - answering questions, writing code, following instructions - is that same continuation, shaped by training on examples of those tasks.

That framing explains most of the behaviour you will run into. The model has no separate store of facts to check against, so a plausible continuation and a true one are equally easy for it to produce. It has no memory between calls beyond what you put in the context window. And it responds to the shape of your input, which is why the same question phrased two ways can get two different answers.

What you can rely on is that the model is very good at pattern completion when the pattern is clear. Give it a worked example of the output you want and it will match the structure closely. That is the single highest-value technique in prompting.