How to Write AI Prompts: Structure Beats Magic Words

Most prompt advice is a list of incantations — "act as an expert", "take a deep breath" — and it ages badly, because a technique extracted from one model's quirks becomes noise in the next. What survives a model change is structure: say what the task is, what the constraints are and what the answer should look like.

Three parts do the work, and you can test them for free on any of the assistants here. Context tells the model what the text is for and who will read it, which is the part people skip and then complain about the tone. Constraint sets length, register, audience and exclusions — "no more than 200 words", "no bullet points", "don't mention price". Answer shape asks for the output you actually need to use: a table with these three columns, a draft email, a list of questions rather than answers. A prompt missing the third part usually gets a wall of prose that has to be rewritten by hand, which is why the shape request saves as much time as the first two.

Iteration is the part that gets sold as a trick but is simply normal: the first answer is a sample, the second is where you correct it, and by the third you have a prompt worth keeping. Two limits make this cheaper over time rather than harder. First, prompt styles age — as models change, instructions that once corrected a specific weakness become redundant, so a prompt that suddenly works less well may need less, not more. Second, long conversations accumulate noise, and "no, I meant…" appended ten times produces a worse result than restating the task cleanly with the context it needs; that also happens to be how you end up with something reusable. All three assistants here — ChatGPT (free, $20/mo Plus), Claude (free, $20/mo Pro, about $17/mo annual) and Gemini (free, Google AI Plus $4.99/mo, AI Pro $19.99/mo) — are enough to practise on, and free tiers are enough to learn on.