How to write AI prompts for content, according to the research
Expert personas, tipping and threats have all been tested at scale and none of them work. Formatting, examples and position do. Here is the difference, with prompts you can paste.
8 minute read
Most prompt advice has never been tested
The advice circulating about prompting is largely folklore that spread because it sounded plausible, and a fair amount of it has now been measured and found to do nothing. That is worth knowing before you memorise another list of magic phrases. It also cuts the other way: a few unglamorous things about how you lay a prompt out have real, replicated effects. The useful split is not between clever prompts and simple prompts, it is between instructions that change what the model has to work with and incantations that only change how the request feels to you. Everything below is sorted on that basis, with the tests that produced the finding named, because in this area you should not take anyone's word for it, including ours.
The expert persona does nothing
Telling a model it is a world-class copywriter is the single most repeated piece of prompt advice, and it has been tested twice at scale with the same result. Mingqian Zheng, Jiaxin Pei, Lajanugen Logeswaran, Moontae Lee and David Jurgens, publishing in Findings of EMNLP 2024, evaluated 162 roles across 2,410 factual questions and four model families, and found that adding a persona to the system prompt produced no improvement over no persona at all, with small negative effects in places, and that predicting which persona would help was no better than random. Wharton's Generative AI Labs reached the same conclusion in its fourth prompting report, testing domain-matched expert personas on GPQA Diamond and MMLU-Pro: no lift even when the expertise matched the question, and less knowledgeable roles often made accuracy worse. Drop the costume. Describe the task instead.
Tipping, threatening and pleading do not work either
Offering a model money and threatening it both circulate as productivity tips, the latter endorsed publicly by a Google founder. Lennart Meincke, Ethan Mollick, Lilach Mollick and Dan Shapiro tested them properly in August 2025 across five models, running each question 25 separate times, which came to 4,950 runs per prompt per model on the GPQA Diamond benchmark alone. They found no meaningful overall performance improvement from either. Individual questions swung dramatically, up to 36 points better or 35 points worse, but unpredictably, so you cannot know in advance whether a given trick will help the specific thing you are asking. One model showed a roughly ten-point gain from one emotional prompt, which the authors called a model-specific quirk rather than a strategy.
Politeness: two studies, opposite answers
This is the one place where the honest answer is that nobody knows. Ziqi Yin and colleagues ran a cross-lingual study across English, Chinese and Japanese and concluded that impolite prompts tend to degrade performance while excessive politeness buys nothing, with moderate politeness usually best. Then Om Dobariya and Akhil Kumar, in an October 2025 short paper, took 50 questions across maths, science and history, rewrote each into five tone variants for 250 prompts total, ran them through ChatGPT-4o and found the opposite: very rude prompts scored 84.8% against 80.8% for very polite ones. Wharton's first prompting report found please versus I order you moved individual questions by up to 60 points in either direction, then cancelled out across the full dataset. Be polite because you want to be. It is not a technique.
Think step by step is no longer free
Chain-of-thought prompting earned its reputation on older models and the ground has moved. The Wharton group's second report tested it on GPQA Diamond and found that for non-reasoning models it lifts average performance a little while increasing variability, meaning it introduces fresh errors on questions the model would otherwise have got right. For models that already reason before answering, the gains were marginal and came with 20% to 80% more time per answer. For content work this matters less than it sounds, because you are rarely asking for a right answer. It matters when you ask a model to pick the best five moments from a transcript: telling it to explain each choice gives you something to audit, which is worth more than any accuracy effect.
What does hold up: structure, position and examples
Melanie Sclar and co-authors showed at ICLR 2024 that models are extremely sensitive to prompt formatting alone, with performance spreads reaching 76 accuracy points on LLaMA-2-13B from changes that preserve meaning entirely. Wharton's first report found formatted prompts consistently outperformed unformatted ones. None of that means there is one correct format; it means a prompt with the material clearly separated from the instruction beats one long blob. Anthropic's own prompting guidance recommends wrapping distinct parts in tags such as instructions and example so the model does not confuse your reference material with your request. And examples beat adjectives every time. Punchy means nothing. Twenty captions you actually wrote, pasted in with the instruction to match their sentence length and rhythm, means something specific.
Four things every content prompt needs
A prompt that works usually contains the material, the constraint, the output shape and the exclusions. Material is the transcript, the comment dump, the twenty captions. Constraint is the real one: 90 characters, 45 seconds, a LinkedIn audience of agency owners. Output shape is what you want back and in what form, including how many options. Exclusions are the tells you refuse to publish. A working caption prompt reads roughly like this: here are twenty captions I wrote that performed well, match their sentence length and their habit of opening with a concrete noun, now give me fifteen variants of this line, every one under 90 characters, no emoji, no rhetorical questions, no the word unlock. Ask for fifteen and choose one. Asking for one and accepting it is where the generic register creeps in.
Three prompts for the jobs you actually do
For clip selection: here is a transcript with timestamps, find the eight moments where the speaker says something that still makes sense with no setup, quote the exact sentences verbatim with their timestamps, add one line on why each stands alone, do not paraphrase and do not invent a timestamp. For comment mining: below are 300 comments from my last video, group them into recurring questions ranked by frequency, paste three verbatim comments under each group as evidence, and delete any group you cannot support with three real comments. Both share the same trick — demanding verbatim quotes gives you an audit trail, so a fabricated theme is obvious in seconds. The same discipline applies once the clips exist: automation like FrameOS handles the cutting, reframing and captioning, and the choice of what is worth posting stays yours.
FAQ
Does telling AI to act as an expert improve the output?
No, and this has been tested twice at scale. An EMNLP Findings 2024 study covering 162 roles, 2,410 questions and four model families found personas produced no improvement over no persona, with small negative effects in places. Wharton's Generative AI Labs found the same for domain-matched expert personas on GPQA Diamond and MMLU-Pro, and that less knowledgeable roles often hurt accuracy. Describe the task instead.
Should I be polite to AI, or offer it a tip?
Tips and threats have been tested and do nothing meaningful. Wharton researchers ran five models with 25 trials per question and found no overall improvement from either. Politeness is genuinely unsettled: one cross-lingual study found impolite prompts degrade output, while a 2025 short paper found rude prompts scoring slightly higher on ChatGPT-4o. Neither effect is large or reliable enough to treat as a technique.
What actually makes a prompt better?
Three things with evidence behind them. Structure: Sclar and co-authors found formatting changes alone can move performance by up to 76 accuracy points, so separate your material from your instruction clearly. Examples: pasting your own writing in beats describing the tone you want. And volume of options: ask for fifteen variants and choose, rather than asking for one and accepting it.
How do I stop AI inventing things in my content prompts?
Demand verbatim evidence. Ask it to quote the exact sentences from your transcript with timestamps, or to paste three real comments under every theme it claims to have found, and to delete any group it cannot support. A fabricated theme with no quotes underneath is visible in seconds. Never let a model be the origin of a statistic you have not checked on the publisher's own page.
Sources
- Technical Report: I'll pay you or I'll kill you — but will you care? · Wharton Generative AI Labs
- Technical Report: Prompt Engineering is Complicated and Contingent · Wharton Generative AI Labs
- When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models · ACL Anthology
- Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design · arXiv
- Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy · arXiv
Related pages
Keep reading
More AI →The best AI tools for social media, by the job you need done
Most AI tool lists are directories. This one is organised by job, because the useful question is not which tool is best overall, it is which step of your week you want to stop doing by hand.
9 minute readThe best AI transcription tools, judged on measured accuracy
On clean read speech almost every engine looks excellent. On meetings, accents and crosstalk the same models lose several points of accuracy — and the published numbers show exactly where.
8 minute readThe best AI writing tools for creators, and what each one is actually for
Most roundups list forty tools that all do the same thing. The useful split is between general assistants, the editing layer, and tools that write from material you already have.
8 minute readTurn one long video into a week of posts
300 credits for 3 days · no card.
Start free