Prompting

What Prompting means

Prompting is the craft of writing the input to a language model so that the answer we want becomes the most likely way for the text to continue. The model does not follow our instruction; it carries on from it, so a prompt that opens like an expert's memo tends to continue like one. Owning this changes how we write to the tools. We stop describing what we want and start setting up the page so that what we want is what comes next: the role, the example, the format shown rather than requested. Before: ask nicely and hope. After: build the opening that leads where we are going, and treat a bad answer as a bad opening, not a stubborn machine.

Not to be confused with

Prompting is not fine tuning. Fine tuning changes the model by training it further on new examples. Prompting changes nothing in the model; it changes what the model is asked, and so which of its answers is most likely.

Prompting examples: where it shows up

• At work: the colleague whose prompts always seem to work is not being cleverer; they show the tool an example of the answer before asking for one.
• At home: the meal plan that came back vague was answering a vague opening, and one line about who is eating and what is in the fridge changes it.
• In the news: stories about people tricking a model into misbehaving are stories about an opening the model continued, because continuing is all it does.

Where it sits

Concept 5 of 5 in Working with AI, a journey in the course Understand AI.

See where Prompting sits on the sky →

Nearest on the sky

• Attention
• The Cooperative Principle
• The Generation Loop
• Hallucination
• The Training Pipeline
• Capabilities and Limits

A filled dot marks a star that is already open to walk.