Part three of the series AI Content: 10 Strategies That Work. A new part every Friday.
A prompt is a brief. Anyone who can brief a person can brief a language model, because the same thing applies to both: the more specific the brief, the less room there is for an average result.
The difference between a bad brief and a good one
A bad brief reads: write an article about investing in cryptocurrencies. The model answers with whatever is statistically most common, which means a generic text that already exists online in thousands of copies.
A better brief looks like this: write a guide to the first steps in cryptocurrency investing for conservative investors aged 40 to 55 who understand mutual funds and bonds but distrust digital currencies. Explain through analogies with instruments they already know. Address three specific worries: volatility, the risk of losing access to a wallet, and the unclear tax treatment of crypto gains. Do not use the word revolution or any comparison to gold. Length 1,200 words, structure: the problem, three steps, what to do this week.
The second brief is not longer for effect. Every sentence in it removes one way the model could slide into generality. The age and knowledge profile stops the text explaining the basics to someone who already knows them. Banning two specific phrases blocks the two most common clichés in the field. Specifying the structure stops the text ending without a recommendation.
Writing a brief like that takes three minutes. Fixing the generic text produced by the first brief takes half an hour and usually ends with you rewriting it anyway.
Four things that belong in almost every brief
The first is context: who the text is for and why it exists. Without it the model picks the most general possible audience, which is an audience that does not exist.
The second is structure: what the result should look like. If you do not specify the structure, you get the one the model has seen most often, which is usually an introduction full of generalities, five equally long sections and a summary that summarises nothing.
The third is perspective: the angle the topic should be approached from. The same topic written from the point of view of someone selling a service and someone complaining about it are two entirely different texts.
The fourth is prohibitions: what must not appear in the text. This point gets left out most often and has the biggest effect, because language models have strong default tendencies towards certain phrasings and will return to them unless explicitly told not to. It is worth putting at least this much into your template: no openings along the lines of „in today’s world“, no „it is not just about X, it is about Y“ constructions, no vague references to studies without a specific source, and no triplets of adjectives.
Chaining and a prompt library
Splitting the work into steps works better than one large brief. First the outline, then the individual sections once the outline is approved, then a check at the end. At each step the model has a narrower task and less room to drift.
Save the prompts that work. It sounds like trivial advice, but after six months the gap between someone keeping a library of twenty tested briefs and someone writing a brief from scratch every day is enormous. The first has a tool, the second has a hobby.
For every saved prompt, note which model it worked on and when. Portability between models is not a given, and a brief tuned for one model produces markedly worse results on another. After every major model update it pays to go through the library again.
What does not work, however widely it circulates
Advice circulates along the lines of „explain the concept without using the letter e“ as an exercise in creativity under pressure. With language models this does not work reliably.
The reason is technical. Models do not work with letters but with tokens, meaning fragments of words. A brief formulated at the level of individual letters therefore misses how the model sees the text at all. The model will break such an instruction regularly and you will spend more time checking than creating.
This goes double for languages other than English, which break into tokens less efficiently. We will come back to that in more detail in the part on personalisation, where the same property has a direct effect on cost.
Fact-checking: the most important part of the whole chapter
Language models make things up. Not because they are faulty, but because they generate a statistically probable continuation of a text, and a statement that sounds probable does not have to be true.
The dangerous inventions are not the conspicuous ones, they are the convincing ones. A model will not tell you that Prague sits on the coast. It will tell you that according to a 2023 study conversions rose by 34 %, which sounds exactly like a sentence you read somewhere, so you pass it on without hesitating.
How much models make up is harder to measure today than you would expect. In its system card from July 2026, OpenAI reports neither a SimpleQA score nor a LongFact score and limits itself to the qualitative statement that the newest model makes slightly fewer factual errors than its predecessor. The fact that the largest player publishes no hard numbers on factual reliability is itself a piece of information worth keeping in mind when you design an editorial process.
Why quotations are the most treacherous
A textbook example is the line „Never test the depth of the river with both feet“, which spreads across the internet and through AI output as a Warren Buffett quote. There is no primary source for it.
You will not find it in the Berkshire Hathaway shareholder letters or in any documented interview. It appears exclusively on quote aggregators, always without saying when and where Buffett said it. The same saying is meanwhile widely published as an African proverb. A model will happily attribute it to Buffett, because that is how it appears online a thousand times over.
Even a quotation whose attribution is probably correct is worth verifying. Francis Bacon is credited with „A prudent question is one half of wisdom“. The Latin original reads „Prudens interrogatio quasi dimidium sapientiae“, meaning a prudent question is as if half of wisdom. That little word quasi routinely disappears in translation, which makes the statement more absolute than it was.
And even then, not even specialist quotation databases have managed to document a literal primary source. A quotation is simply the format that travels the internet without a source more easily than anything else, and models learn from precisely that internet.
The rule that saves you embarrassment
Verify every number, name, date and quotation from AI output against a primary source before you publish it. Not in a second AI tool, not in the first search result, but in the primary source.
A primary source means the study, not an article about the study. The court decision, not a law firm’s summary. The provider’s price list, not a comparison table on a blog. The gap between those two levels is exactly where the errors that then live for years enter circulation.
The practical routine that works is simple. While writing, attach a link to every claim containing a number, even if only as a note. Then go through the text before publication looking for sentences that assert something measurable and have no link. Either you add the source or you cut the sentence.
It sounds laborious, but it is incomparably cheaper than a single public correction. An error you publish is yours, not the model’s.
Previous part: A publishing plan that assumes volume does not work. Next Friday: Personalisation that is more than a different salutation.
The whole series: AI Content
1. Why most advice about AI content no longer holds · 2. A publishing plan that assumes volume does not work · 3. The prompts that make a difference, and the fact-checking · 4. Personalisation that is more than a different salutation · 5. Visibility in the year people stopped clicking · 6. Where AI ends and a human begins · 7. Images, and how to spot a marketing myth · 8. Measurement that tells you something · 9. One piece of content, several channels, without copy-paste · 10. Law and ethics, specifically, as they stand in August 2026 · 11. Build a system, not a pile of articles · 12. What to take from the whole series