Most disappointing AI answers were earned. Not because the model is weak — because the ask was vague. This level gives you the recipe that fixes it, and the follow-up moves that rescue almost any answer.
The model answers the question it was given — not the one in your head. Every detail you leave out, it fills in with the most average assumption possible.
Remember Level 01: it predicts the most likely continuation. Ask an average question and the most likely continuation is an average answer — the one that would fit anyone. The fix isn't magic words. It's giving it the details that make you not-average.
"Help me write a CV."
"You're a recruiter who screens hundreds of CVs for junior retail roles. Rewrite my CV below for a shop-floor job at a sports store. I'm 17 with no formal work experience — but I've refereed football and run a sneaker resale page. One page, punchy bullet points."
"Explain photosynthesis."
"You're a patient tutor. Explain photosynthesis for a Year 10 exam — the equation, the two stages, and the three things examiners always ask. Then quiz me with 3 questions, one at a time."
Same model. Same effort to type. Completely different result — because the strong version answers your situation, not the average of everyone's.
Three parts, every time: who it should be, exactly what you want (with your real details), and what shape the answer should take.
Role points the model at the right slice of its patterns — tone, expertise, point of view. Task specifics are steering: every concrete detail ("Year 10 exam", "$40 budget", "no work experience") cuts away a thousand wrong answers. Format means you get something you can use immediately — a list, a table, questions one at a time — instead of a wall of text.
You've already played with this: the Prompt Lab on the course page builds prompts in exactly this shape. This level is about making it a reflex.
Stopping at the task. "Write a birthday message for my friend" is a task with no material. Who's the friend? What's one real memory? What tone — roast or sweet? The model can't know what you didn't say. If a stranger couldn't do the job from your instructions, neither can the AI.
Nobody gets the perfect answer on message one — and you don't need to. The pros treat the first answer as a draft and push. Two moves cover almost everything:
"Shorter." "Less formal." "More like how a real teenager talks." "Keep point 2, cut the rest, go deeper on it." Every steer teaches it what you actually wanted.
The single most underused prompt in existence:
Now it interviews you, pulls out the details you didn't think to give, and the next answer lands twice as close.
It answers the question it was given. Specifics are steering, not decoration.
Three slots. Fill all three and you're already ahead of most people using these tools daily.
Steer it, or make it ask you questions. The follow-up is where gold happens.
← back to all 10 levels · stuck on anything? ask peter.