Why do people get stuck?
Not because they do not know what they want. Because what they want is in their head, not in sentences. You know it should "read like the ones Julia writes", but you cannot write down what makes them so.
Most of prompt writing is not writing at all, it is excavation: getting out of yourself the things you take for granted. For that the model is a better tool than a blank page, because it can ask questions.
Recipe 1: have it interview you
The most useful of the three. Do not ask for a prompt, ask for an interview.
We are going to write a prompt together. The task: [in one sentence, however you can].
Do not write the prompt yet. First ask me at most three questions at a time about whatever would otherwise leave the task ambiguous. When I answer, ask again, until nothing is unclear. When there is nothing left, say so, and then write the prompt.
"At most three questions" is not decoration. Without it you get a fifteen-point questionnaire nobody fills in. With it you get a conversation, and in three or four rounds things surface that you would never have written down: who it is for, which direction is off limits, what happens when a value is missing.
Recipe 2: have it repair the one you have
For when you have a prompt but the result is not good.
Here is a prompt and an answer that did not come out well.
[your prompt] [what you got] [one or two sentences: what is wrong with it] Do not rewrite the prompt. First list what information is missing from it for the answer to be good. Then propose concrete additions, justifying each sentence.
The restriction is the point here too. If you have it rewrite straight away, you get a longer, more flowery prompt and you will not know what made it better — or whether it did. Ask for the gap list first and you decide what goes in.
Recipe 3: reverse-engineer from a good one
Works when you have a finished piece of work that is exactly what you want more of.
Here is a text that is exactly the kind I want from now on.
[the text] Describe what instruction could have produced this: tone, structure, length, word choice, what it includes and what it avoids. Then turn that into a prompt.
This is the fastest route to describing your own style. You do not have to believe what you get back — but it will name things you never noticed about your own writing. Give it three or four samples and it gets sharper still.
What to watch for
Trim what comes back. The model tends to inflate a prompt: role descriptions, politeness formulas, "very important" qualifiers. None of these improve anything. Go through it and keep what is genuine information — throw the rest out. The five elements in The Anatomy of a Working Prompt make a good filter for this.
Do not accept the first version. Meta-prompting works when it runs two or three rounds. The first prompt is typically generic; the second is already yours.
Do not use it to fix a bad answer. If there is just one specific error in the output, the three sentences in Fixing a Bad Answer are quicker.
Test it, don't just read it. A prompt is good because it produces a good answer, not because it sounds good. Run it on three real cases before you adopt it.
When it does not help
When the material is missing. If the model has nothing to work from, the best prompt in the world will not help — access to the data has to come first. That is the first question in Can AI Do This Task?, and it comes before everything else.
When you cannot say what a good answer is either. If you cannot tell whether an output is acceptable, the prompt will not get better by being more elegantly worded. That is not a prompting problem.
When the task is really two. Meta-prompting tends to build one big prompt that knows everything. If the task has two different outputs, you need two prompts.
When you have a good prompt
Save it. A working prompt is like a well-proven template: writing it twice is a waste, and by the third time you will not remember which version was the good one. That is the subject of Prompt Library, a separate article.
What to take away
Do not ask for a prompt, ask for questions. Meta-prompting works when the model first extracts from you what you take for granted — and only then writes. Trim what comes back, test it, and if it holds up, save it.