Most prompting mistakes come from a wrong picture of what the model is. A large language model (LLM) is a system trained to predict the next piece of text, called a token, given everything that came before it. A token is roughly a word fragment: “prompting” might be two tokens, a space and a comma one each.
When you send a prompt, the model reads the whole conversation so far, picks a likely next token, appends it, and repeats until it decides to stop. Everything it “knows” comes from two places:
- Training data: patterns learned from a very large body of text up to a cut-off date. This is broad but fuzzy, and it can be out of date or simply wrong.
- The context window: the text in front of it right now — your instructions, pasted documents and the conversation history. This is precise, and it is the part you control.
Three practical consequences follow.
- The model fills gaps with the most typical answer. A vague prompt gets a generic response, because generic is statistically likely. Specific context pulls the output toward your situation.
- It can sound confident and be wrong. Fluent text is what it is optimized for, not verified fact. If accuracy matters, give it the source material and ask it to quote from it, then check.
- Output varies between runs. Most tools sample with some randomness. The same prompt can give different answers, which is why “consistent results” is a skill rather than luck.
Try this
Ask a chat assistant: “Write a product description for a water bottle.” Then ask again with context: who the customer is, the three features that matter, the tone, and a 60-word limit. Compare the two answers. The difference is the whole course in miniature: the model did not get smarter; you gave it less room to guess.