1. The rule is buried in the middle
Models pay most attention to the start and the end of a prompt. A rule tucked inside a long paragraph of context is the one most likely to be missed.
Put your rules last, as a short list under a label like “Rules:”. One rule per line.
2. The rule can't be checked
“Keep it short” means different things to different readers. So does “be professional”.
Make every rule something you could check yourself: “under 120 words”, “3 bullet points, no more”, “no exclamation marks”, “write at a 10-year-old's reading level”.
3. Two rules fight each other
“Be detailed” and “keep it under 50 words” can't both win. When rules conflict, the model picks one, and it may not be yours.
Read your rules once as a set. If two pull in different directions, say which one matters more.
4. The chat got too long
In a long conversation, early instructions get less weight than recent messages. That's why ChatGPT follows your format for ten replies, then drifts.
Repeat the key rule in your latest message, or start a fresh chat with the rules at the top. For rules you always want, use Custom Instructions so they apply to every chat.
5. You said what not to do, but not what to do
“Don't use jargon” leaves a gap. “Use everyday words a customer would use” fills it.
Pair each “don't” with a “do”. Models follow a positive target more reliably than a ban.
6. The task is doing too many things
Ask for research, a summary, a rewrite and a table in one go, and something gets dropped.
Split it: one prompt per step, or a numbered list of steps with the output you want from each.
7. There's no example
If the format matters, show it. One short example of the output you want beats a paragraph describing it.