The prompts that circulate most widely are the ones that are most fun to share. That is not the same as the ones that work, and the gap between them explains a lot of disappointing afternoons.
If you have collected a folder of "magic prompts" and found that your results did not obviously improve, nothing is wrong with you. Most of what is shared is decoration on the outside of the request, and the thing that actually determines the answer is on the inside.
The tricks, and what they actually do
"Act as a world-class expert in…"
This was genuinely useful with the models of a few years ago, which had a habit of answering at the level of the question. Current models already aim high by default. What the persona still changes is register — vocabulary, formality, what gets assumed rather than explained — which is a real effect and a narrow one. Telling it to answer as a tax accountant will change the tone. It will not give it access to facts about your tax situation that you did not provide.
Keep personas when you want a particular voice. Stop expecting them to add expertise.
"I'll tip you $200" and "my career depends on this"
These came out of a period when people were testing emotional framing and reporting striking results. They have not held up as reliable techniques across models and tasks, and the reason is worth understanding: they were pushing on how much effort the model put in, which the current generation largely does by default. There is no reason to think you are being punished for not offering a bribe.
"Take a deep breath and work through this step by step"
The interesting half of this one is real. Asking for reasoning to be worked through rather than jumped to does help on problems with several dependent steps — arithmetic, logic, anything where an early mistake propagates. The deep breath is decoration. "Work through this one step at a time and show your reasoning" is the part doing the work.
The 500-word preamble
Long ceremonial framing — rules, personas, promises, formatting demands, all before the actual question — dilutes the instruction that matters. The request should be the biggest thing in the prompt.
The five things that genuinely change the output
All of them are unglamorous, and all of them are about supplying information the model does not have.
1. Context: who and what for
"Write a product description" and "write a product description for a £40 hand-thrown mug, sold to people buying a wedding present, on a page where the photographs already show what it looks like" are different requests. The second cannot produce generic output, because generic output would not fit.
2. Constraints, including the negative ones
Length, audience, reading level, what to leave out. Saying what you do not want is the single most underused instruction available: no bullet points, no marketing language, do not restate the question, do not hedge. These work, and they are specific enough to be checkable.
3. An example
One sample of the thing you want carries more information than a paragraph of adjectives describing it. Paste a previous email you were happy with and ask for another in that style. Style is very hard to specify and very easy to demonstrate — and if you have no example, an inversion works nearly as well: paste something in the style you are trying to avoid.
4. The shape of the answer
Say what you want back: a table with these three columns, five options with one line each, a paragraph under a hundred words, a checklist. Vague requests produce essays, because an essay is the safest answer to an unspecified question.
5. Making it ask you first
The highest-value sentence in everyday use is some version of: before you write anything, ask me up to five questions that would change your answer.
This inverts the usual failure. Instead of the model guessing at everything you left out and you discovering the guesses in the draft, it tells you which gaps matter — and its questions are frequently ones you had not thought to consider. For anything longer than a paragraph, this beats every clever phrasing on the internet.
What this looks like in practice
Before:
Act as an expert career coach with 20 years of experience. Write me a professional cover letter for a marketing job. Make it compelling.
After:
I'm applying for a marketing coordinator role at a 30-person software company. Here's the job posting, and here's my CV. Write a cover letter of about 250 words. It should open with why this specific company rather than a generic opener, cite exactly two things from my CV that match the posting, and sound like a person, not a template. No "I am writing to express my interest". Ask me anything you need before you start.
The second is longer, and none of the extra length is technique. It is information: the role, the company size, the source material, the word count, the structure, a banned phrase, and permission to ask. The first prompt could have been sent by anybody about any job, which is precisely why it comes back reading that way.
A skeleton worth reusing
- The task, in one plain sentence, first.
- The context the model cannot see: who it is for, what you have already tried, what happened.
- The material, pasted in — the document, the data, the previous version.
- The constraints: length, format, tone, what to avoid.
- An example of good, or of bad.
- "Ask me any questions first."
Not every request needs all six. A short one needs one and four.
The part nobody shares
Good results come far more often from the second and third message than from a perfect first one. The people who consistently get useful output are not writing better opening prompts; they are reading the reply, saying precisely what is wrong with it, and asking again.
"Too long, and the third paragraph is a cliché — cut it and make the ending concrete" is worth more than any prompt template, because it is information the model genuinely did not have until you looked at the draft. That is not a workaround for a weakness. It is what using the tool actually consists of, and it is the reason a folder of magic prompts never quite delivers on the promise.