Anatomy of a good prompt
Collections of prompts to copy go stale in a month. One framework, learned once, improves every prompt you write after it.
There is a whole genre of content that is just lists of prompts to paste. It is popular because it feels like a shortcut, and it is close to useless, because the prompt that works is the one carrying your situation, and a stranger cannot write that for you.
So learn the shape instead. Four parts, and you can hold them in your head.
Role, Task, Context, Format
is who it should be for this task. "You are a media planner pitching a sceptical CMO" produces different work from "you are a copywriter". It sets the lens and, just as importantly, the level. Ask for a senior perspective and you get one.Role
is the actual ask, with a real verb. Summarise, draft, compare, critique, rank, rewrite. Vague verbs like "help me with" or "look at" produce vague output, because you have not said what done looks like.Task
is the background: the audience, the constraints, the inputs, what good looks like, what to avoid. This is the part almost everybody skips, and it is the part that does most of the work. If you only improve one thing after reading this chapter, improve this.Context
is the shape you want back. A table with these columns. Five bullets, no preamble. Two options, each under 60 words. Without it you get an essay, and then you spend your time reformatting instead of thinking.Format
Act as . the following, for . Context: . Return it as .[role][Task verb][audience][what matters, what to avoid, what good looks like][format]
What the difference actually looks like
Take a real job: social captions for a bank's Ramadan campaign.
Write some social media captions for a bank for Ramadan.
That returns exactly what you would expect. Generic warmth, crescent moons, the word "blessed" several times, nothing you could put in front of a client.
Act as a regional social copywriter who has worked on banks in the Gulf. Draft caption options for a retail bank's Ramadan campaign, aimed at working adults aged 25 to 40 in the UAE.
Context: the campaign is about a savings product, so the tone is practical and warm rather than devotional. The client is conservative and will reject anything that sounds like it is monetising the holy month. Avoid crescent-moon cliches, avoid the word "blessed", and do not use religious imagery to sell a financial product. Arabic and English both matter, so keep the English simple enough to translate cleanly.
Return eight options in a table: caption, tone in one word, and the risk you see in it.
Same model, same minute, completely different output. Nothing in the second version is clever. It is just the brief, written down. Most of what people call prompt engineering is really just refusing to leave the brief in your head.
The part everyone gets wrong
People treat context as background information. It is not. The most valuable context you can give is : what the client will reject, the phrase that always gets cut, the angle that has been tried twice and failed.negative
A model does not know your account's history. Left to itself it will produce the most statistically ordinary version of the thing, which is exactly the version that gets marked up in review. Telling it what not to do removes more bad output than any amount of telling it what to do.
When something comes back wrong, resist the urge to rewrite the whole prompt. Ask instead which of the four parts you left out. It is almost always Context, and usually the negative half of it.
Where this is going
You will notice that a lot of the context you type is the same every time. The same client, the same brand rules, the same list of things that always get rejected. Retyping that on every prompt is the obvious next thing to fix, and that is exactly what the next chapter is about.
Take the last prompt you wrote that gave you a disappointing answer. Do not rewrite it. Just add the context you left out, especially what you did not want, and run it again. Keep both answers side by side.