What goes into a prompt that works?
A prompt that works describes the job the way you would brief a capable new staff member on their first morning. Say who you are and who the customer is. Give the facts the model cannot know: the price, the suburb, the hours, what is included. Say what you want back, how long it should be and in what tone, and if you have an example you liked, paste it in.
The model makers' own guides say much the same. OpenAI's prompt engineering guide, Anthropic's prompt engineering documentation and Google's prompting guide for Gemini all recommend clear instructions, background and examples. None of them recommends magic words. A line like "act as a world-class marketer" does far less than the one fact you forgot to include.
What changes when a surf school asks for the same ad twice? (example figures)
A surf school in Bunbury wants a Facebook ad for its school holiday group lessons, with $300 to spend over the two weeks. The first prompt is one line: "Write a Facebook ad for my surf school." The second reads: "I run a surf school in Bunbury, WA. Write a Facebook ad for our school holiday group lessons: 2 hours, $65 a person, boards and wetsuits included, ages 8 and up, groups of 6 with one instructor. The readers are parents within 30 minutes' drive. Under 60 words, plain and friendly, no exclamation marks. End with: Book at the link."
| What came back | The one-line prompt | The full brief |
|---|---|---|
| Who it speaks to | Anyone who has ever thought about surfing | Parents near Bunbury in the holidays |
| Price | Missing | $65, as given |
| Age and group size | Missing | 8 and up, groups of 6 |
| What to bring | "All gear provided", a guess | Boards and wetsuits included |
| Next step | A generic "Contact us today" | Book at the link |
| Editing needed | A rewrite | A word or two |
Both versions cost the same $300 to run. Only the second answers a parent's questions before they tap, and the time it took to write was about two minutes.
Is there a prompt behind the prompt?
Yes. Almost every AI app sends its own instructions ahead of yours, called a system prompt, which sets the tone, the rules and what the app will refuse. You never see it, and it is the reason the same underlying model sounds chatty in one app and clipped in another.
It is also why text from elsewhere can cause trouble. A web page or an email the AI reads can contain lines written to look like instructions, and the model may treat them as if they came from you. That problem has its own name, prompt injection.
Why can a longer prompt give a worse answer?
Because length and clarity are different things. A prompt that contradicts itself (formal, but chatty), buries the request under three paragraphs of background, or pastes a long document with the question in the middle leaves the model guessing which part counts. Models also pay most attention to the start and end of a long input, as the entry on context windows explains. Put the request first or last, cut what the model does not need, and when an answer misses, say what was wrong instead of starting from scratch.
How do you brief Cora?
With CoraVision the briefing is a conversation: you talk to Cora out loud, the way you would brief a person across a table. The onboarding call takes about 25 minutes and needs no preparation. Cora asks the questions a good prompt would answer, starting with your goal in your own words and your average sale, and works out your lifetime value from what you say. Your figures need no typing in either, because once Google Ads, Meta and Google Analytics 4 are connected, CoraVision reads them every night.
Questions
Do I need to learn prompt engineering?
For everyday use, no. Writing the brief you would give a capable new employee gets you most of the way. Prompt engineering is the name for doing this carefully inside software, where one prompt runs thousands of times and small changes in wording get tested.
Can I reuse a prompt that worked?
Yes, and it saves time. Keep the ones that worked in a document, with blanks for the parts that change, such as the product or the date. Models get updated, so check the output each time rather than trusting last month's result.
Is a spoken request a prompt too?
Yes. A voice request is turned into input the model can read, and it does the same job as typed text. Speaking often makes it easier to include the background, because you can talk it through the way you would with a person.