You're steering an autocomplete

Two people ask the same AI for help. One gets a generic blob of filler; the other gets exactly the email, plan, or code they needed. Same model, same day, same subscription. The difference is entirely in the prompt — and once you see why, writing good ones stops being guesswork.

The key is remembering what the machine actually does. As the How ChatGPT Actually Works course shows, a language model doesn't "understand your request" and then go fulfill it. It reads your text and predicts a plausible next token, over and over. Your prompt isn't a wish submitted to a genie — it's the beginning of a document the model is finishing.

Scroll to watch the same model handle a vague prompt and a specific one.

The model completes — it doesn't read minds

Every detail you leave out of a prompt doesn't disappear. The model still has to pick some audience, some tone, some length — and with nothing to go on, it picks the statistically average one. That's why lazy prompts produce that unmistakable beige, "AI-sounding" text: you asked for the average of the internet, and you got it.

Steering, not spellcasting

This mechanism also tells you what prompting is not. There are no secret magic words, no incantation that unlocks a hidden smart mode. A prompt works when it gives the model information — about the task, the reader, the format, the constraints — that narrows down what a good completion looks like.

That reframing turns prompting from folklore into engineering:

• The model will do something with whatever you give it — the question is whether you left the important choices to chance.
• More relevant context in the prompt → sharper, more useful probabilities out.
• Bad output usually means an underspecified prompt, not a broken model.

The rest of this course is that idea, applied five ways. First up: the single highest-leverage habit — saying what you actually want.