The four parts every good prompt has: task, context, role and output. With a weak and a strong example side by side.
You open ChatGPT, type "write a LinkedIn post about our new service", and get something you could have found on any company page. Then you wonder whether AI is overrated, or whether you are doing it wrong. Usually it is the second one, and the fix is smaller than it feels.
An AI marketing prompt is the brief you give a language model such as Claude, ChatGPT or Gemini to get a marketing result: a post, an email, a plan, an analysis. A good one contains four things: the task, the context, the role and the output format. Leave one out and the model fills the gap with its own guess, which is where generic copy comes from.
In one line: an AI marketing prompt is the briefing you give a language model, made of a task, context, a role and an output format, and it is only as good as the context you put in it.
A prompt is one briefing. You write it, the AI answers, done. The quality of the answer follows the quality of the briefing, because a language model does not understand your intention. It predicts text from patterns, so the clearer and more structured your prompt, the more relevant the result.
Two things follow from that. A one-line prompt is a bad prompt, because a single sentence gives the model almost nothing to predict from. And a prompt only exists inside that one conversation: in the next chat you write it again, or you paste it from your own prompt collection. That second point is why a saved, well-structured prompt earns its place in your week, and why mayday keeps a curated set of them in The Collection rather than a pile of one-liners.
A good prompt contains a task, context, a role and an output format. Google's prompting guide for Gemini names the same four areas as persona, task, context and format (Google Workspace, writing effective prompts), so this is not a mayday invention, it is how the tools are designed to be used.
Three of the four are quick. Context is the one that takes effort, and it is the one most people skip.
Tell the model exactly what to produce, in a size it can produce well. "Write a 150-word LinkedIn post announcing our new onboarding workshop, for marketing managers at B2B companies" is a task. "Do my LinkedIn" is a wish. Anthropic's own guidance puts it simply: show your prompt to a colleague with minimal context on the task and ask them to follow it; if they would be confused, the model will be too (Anthropic, be clear and direct).
Context is the part that turns a generic answer into yours. It covers who you are or what your business does, what you stand for, who your target audience is, what the situation is right now, and what a good result looks like in your eyes. On client work, the pattern mayday sees most often is the opposite: people give almost no context and expect the model to already hold everything that is in their head. It does not. A language model knows the average of the internet, and nothing about your clients, your tone or your last campaign unless you tell it.
The practical fix is a small knowledge base: a short document about your business, your audience and your voice that you paste in, attach, or keep in a project folder. Google's tip number five is the same idea in their words: personalise the output with information from your own files (Google Workspace). Write it once and every prompt after that starts from a much higher floor.
"You are a marketer" is too vague to change anything, because the model already behaves like a generic marketer by default. Give it a specialism and the level you want the answer at: "You are a social media marketer specialised in LinkedIn and personal branding, writing for a senior audience that dislikes hype." The role decides which patterns the model reaches for, so a precise role gives you a precise answer.
Length, format, tone, language, structure. If you do not describe the output, you get whatever the model decides to invent, and it tends to invent bullet lists, headers and a closing call to action you did not ask for. "Return one post, under 150 words, no hashtags, no emojis, sentence case, end on a question" leaves nothing to chance.
A good prompt has to be complete, which is different from long. Google reported that the most successful prompts in its Workspace Labs programme averaged around 21 words (9to5Google on Google's prompting guide). That number surprises people who have been told to write mega-prompts, and it deserves an honest reading: 21 words is already three or four times the one-liner most people type, and in Google's own method the heavy context lives in attached documents, not in the prompt itself.
Pro tip: keep the prompt short, keep the context in a file. Write your business, audience and voice down once in a knowledge base document, then attach or paste it next to a tight prompt. Short prompt, rich context, every time.
So the working rule is this: keep the prompt tight, keep the context in a document you reuse. The prompt says what to do and how the result should look. The knowledge base says who you are. Together they are complete; apart, neither is enough.
Weak vs strong, part by part
Task. Weak: "Write a post about our new course." Strong: "Write one LinkedIn post announcing our new 'AI for marketing teams' course, aimed at in-house marketing leads at companies of 20 to 200 people."
Context. Weak: none. Strong: "Attached: our brand file and two past posts that performed well, with a note on why. The course is practical, four weeks, no coding. The audience is curious about AI but tired of hype."
Role. Weak: none, so a generic marketer. Strong: "You are a B2B content marketer specialised in LinkedIn, writing in a calm, direct voice."
Output. Weak: none, so whatever the model decides. Strong: "Under 150 words. No hashtags, no emojis, sentence case. Open with a concrete situation, close with one question."
The weak version is faster to type. The strong version is faster to use, because the first result is close enough to publish after one edit instead of five rewrites.
mayday tests prompts on real client work before they go into The Collection, and the four-part structure above is the first thing each one is checked against. Every prompt there carries its task, context slots, role and output format already written, so you fill in your own business and run it in Claude, ChatGPT or Gemini, whichever you already use. Join mayday if you would rather start from a prompt that is already complete.
A good prompt is a complete briefing: it tells the model what to do, what to know, who to be and what to hand back.
What is an AI prompt?
An AI prompt is the written instruction you give a language model such as Claude, ChatGPT or Gemini to get a result. In marketing, that result is usually copy, a plan or an analysis, and the quality of the result follows the quality of the prompt.
What are the four parts of a good prompt?
Task (what to do), context (what the model has to know about you, your audience and the situation), role (which expert it should behave as) and output (the format, length and tone of the result).
How long should an AI prompt be?
As long as it takes to be complete. Google found successful prompts in its Workspace programme averaged around 21 words, with the detailed context kept in attached documents. A one-line prompt with no context is almost always too short.
Why does my AI copy sound generic?
Because the model has to guess what you left out, and it guesses the average. Add context about your business and audience, a specific role, and a clear output format, and the guessing stops. More on that in why your AI marketing copy sounds generic.
Does a prompt remember my context next time?
No. A prompt lives inside one conversation. In a new chat you paste it again, together with your knowledge base. Reusable instructions that the model applies automatically are a different thing, usually called a skill or a custom instruction. More on that in prompts, skills, workflows and agents.
Check any skill before you install it. What it does, who it's for, and whether it's safe.
