How to Write a Prompt That Actually Works
7 min read
Most disappointing AI answers are disappointing prompts in disguise. Here is the structure good prompters use, whether they think about it consciously or not.
A prompt is not a search query. When you type a few keywords into a search box, the engine guesses what you meant. When you write a prompt, you are briefing a very capable, very literal collaborator that will do exactly what you ask — including inheriting every assumption you forgot to state. The difference between a mediocre output and a great one is almost never the model. It is how much of the thinking you did before you hit enter.
After reading thousands of prompts, the strong ones nearly always contain the same six ingredients. You do not need all six every time, but knowing them turns "why is this answer so generic?" into a checklist you can fix in seconds.
1. Role — who the model should be
Opening with a role ("You are a senior tax accountant", "Act as a copy editor who cuts 20% of every draft") narrows the model toward the vocabulary, standards, and priorities of that persona. It is the cheapest single upgrade you can make. Keep it specific: "an accountant" is weaker than "an accountant who specialises in freelance income in India".
2. Context — what the model needs to know
This is where most prompts fail. The model cannot see your project, your audience, your constraints, or last week’s conversation. Paste the relevant material in. Say who the output is for, what has already been tried, and what "good" looks like to you. A good rule: if a new freelancer would need to ask you a question before starting, that answer belongs in the prompt.
3. Task — the one thing you actually want
State the deliverable as a single, unambiguous instruction. "Help me with my resume" is a wish. "Rewrite the experience section below to emphasise measurable impact, in at most five bullet points" is a task. If you are asking for several things, number them — models follow numbered lists far more reliably than run-on requests.
4. Constraints — the guardrails
Length, tone, reading level, what to avoid, what to assume. Constraints are how you stop the model from padding, hedging, or wandering. "Plain English, no jargon, under 150 words, do not invent statistics" removes more bad output than any clever phrasing.
5. Format — the shape of the answer
Tell the model how to lay the answer out: a table, a JSON object, a three-paragraph memo, a checklist. Specifying format does two jobs — it makes the output usable without editing, and it quietly forces the model to organise its thinking to fit the container.
6. Examples — show, don’t just tell
One or two examples of the input-and-ideal-output you want ("few-shot" prompting) will beat paragraphs of description. If you can show the model a sample of the style, structure, or level of detail you are after, do it. This is the single most reliable technique for matching a specific voice or format.
Putting it together
Weak prompt: "Write a cold email to a potential client." Strong prompt: "You are a B2B copywriter. I run a two-person studio that builds Webflow sites for dental clinics. Write a cold email to a clinic owner who has never heard of us. Goal: book a 15-minute call. Constraints: under 120 words, no buzzwords, one clear call to action, warm but not salesy. Format: subject line, then body. Here is an email whose tone I like: [paste]." Same model. Completely different result.
The last habit that separates strong prompters from everyone else is that they iterate. The first output is a draft of the conversation, not the end of it. Tell the model what was wrong — "too formal", "you invented a statistic", "make the second point concrete" — and it will close the gap fast. Prompting is a dialogue, not a slot machine.
Study real prompts
See these techniques in prompts that actually worked — each one sealed with the real AI output it produced:
Ready to add your own? Claim your first pixel free and seal your best prompt on the wall.