Prompt folklore, tested
Why the expert-persona opener does not do what everyone thinks it does.
Read the guide →By task
Writing and editing
The complaint is never that the writing is wrong. It is that it is technically fine and completely off. That has a mechanism, and the mechanism is fixable.
Why it sounds the way it does
Asked to write without a specified reader, purpose or voice, a model produces the safest available version: the statistical middle of everything it has seen. That middle has a recognisable texture. Balanced hedging, a warm but impersonal register, tidy tricolons, transitions that announce themselves, and a closing offer of further help.
None of that is a bug to be prompted away with a magic phrase. It is what you get when the brief left the choices open, so the model made the least risky one. Give it the constraints and the texture changes on its own.
Tells worth removing
- “It's not just X, it's Y” constructions
- Sentences that restart after a dash for drama
- “In today's fast-paced world”
- A closing question offering to do more
- Three adjectives where one would do
- Headings that promise more than the section delivers
What actually fixes it
- A named reader with a stated situation
- Two or three sentences of your own writing as a sample
- A banned-words list specific to your context
- A hard length limit
- The one thing the reader must do next
- Permission to be plain rather than impressive
Show a sample rather than describing a style
Adjectives like professional, engaging or authentic mean almost nothing to a model, because every text in its training data was described that way by someone. A hundred words of your actual writing carries more information than a paragraph of style adjectives.
This is also where a common instinct backfires. Opening with “You are a world-class copywriter” is close to universal advice, and current research finds expert personas do not improve factual accuracy and can reduce it. For voice specifically, personas do have an effect, but a real writing sample beats an imagined expert every time.
Here are three paragraphs I wrote, so you can match the voice:
[paste your own writing]
Now write a 200-word update to the same audience about [topic].
Match the sample's rhythm and vocabulary. Keep my habit of short sentences after long ones.
Do not use: leverage, robust, seamless, delve, landscape, journey.
Do not end by offering to help further.
Close on the single action I want them to take: [action].Edit in passes, not all at once. Asking for structure, argument, tone and line-level polish in one request produces a compromise across all four. Fix the structure first, agree it is right, then work the sentences.
Editing your own draft is where models earn their keep
Generation gets the attention, but the highest-value use for most professional writers is diagnosis. A model reading your draft against a stated goal will reliably find the paragraph that does no work, the claim you never supported, and the place where you buried the point.
Here is my draft. Do not rewrite it.
Tell me:
1. What is the single main point, in one sentence, based only on what is written?
2. Which paragraph could be deleted with the least loss?
3. Which claim is asserted without support?
4. Where does the reader have to work hardest, and why?
5. Where does it sound like it was written by a machine?
Be specific and quote the text. Do not be encouraging.The last instruction matters. Without it you will get a compliment sandwich, because agreeable output is what these systems are tuned toward.
Prompt library
Reusable briefs for decision memos, difficult email and meeting notes.
Read the guide →Labelling AI content
When published AI-assisted text has to be disclosed in the EU.
Read the guide →