Open this lesson in your favourite AI. It'll walk you through the why, explain the demo, and quiz you on the try-it list.
If you don't have even a rough mental model of what a language model is doing with your prompt, every trick you learn feels like superstition. A prompt is just the left-hand side of a very expensive autocomplete — that's the whole magic.
Given "The capital of France is", the model picks the next token with the highest probability given everything you wrote. "Paris" wins because the training data makes it win.
Your prompt is a narrowing funnel. Every word shifts the probability distribution over the next token. "Write a poem" pushes toward poem-shaped continuations. "Write a poem, rhymed ABAB, about loss" pushes harder. "Write a 4-line poem, rhymed ABAB, about loss, in Rilke's voice" pushes harder still.
Bad prompts under-specify the funnel. Good prompts specify it until only good completions fit.
Paste the sentence "Monday, I" into your AI chat and ask it to continue in exactly 10 words. Now try "Monday, I woke up feeling". Then "Monday morning, in Berlin, I woke up feeling". Watch the output collapse from generic to specific as you add constraints.
I want to see how prompt specificity affects outputs.
Complete each of these in exactly 10 words, then in one sentence explain how adding words changed the output distribution:
1. "Monday, I"
2. "Monday, I woke up feeling"
3. "Monday morning, in Berlin, I woke up feeling"