How LLMs Work, and How to Ask So They Stop Guessing
A model predicts the next token from its context window and nothing else. That explains confident wrong answers and the fixes: put the source in the window, give it an out, and write the prompt as a spec that answers five questions.