What this technique does
To stop ChatGPT from making things up, the line to write is not “be accurate” but what to do when it runs into something it does not know. Constraints are the boundary of the result. Settle the don’ts, the length, the tone and the phrases you will not use in advance, and there is less to delete afterwards.
Pairing a ban with an alternative gets it followed far better than the ban alone. “If it is not in the source, write ‘not in the source’” beats “do not guess”. Once the option of leaving a blank is closed off, the AI fills it with a plausible value.
Good example / bad example
Bad
write a blurb for our product. make it accurateGood
Please write a product blurb using only the product sheet below as your source.
Do not use figures or awards that are not in the sheet. If there is no evidence for a sentence, drop it, and collect what you dropped under "needs checking" at the end. Around 150 words, in a plain descriptive voice with no hype.The first answer mixes in phrases like “best in class” that you cannot verify anywhere, along with plausible-looking numbers. The second comes back with the unsupported sentences removed and a list of what went missing, which makes fact-checking easy.
How to write it
- Do not write anything that is not in the source below; where you do not know, write “not in the source”.
- Write around {length}, in a {tone} voice.
- Do not use {phrase or word}.
Watch out
Line up ten constraints and the earlier ones start to blur first. Keep the three or four that really have to hold and add the rest once you have seen a result.
Stack up nothing but bans and the writing gets short and flat. Alongside the don’ts, put one line about what to do instead.
Unfamiliar terms? See Aha AI: hallucination