An ignored rule never looks expensive when it happens. It looks expensive a week later.
Take a made-up five-person marketing agency. The account manager has a standing rule in ChatGPT: "Never quote a price. Prices come from the rate card." On a busy Thursday, a draft goes out that says a landing page "usually runs about $1,500." The real rate is $2,400. Now someone has to walk that back with a client, and the conversation costs more than the rule ever saved.
When I set up ChatGPT for small businesses, most of the work isn't writing clever prompts. It's making the rules hold. And when a rule gets skipped, it's almost always one of six causes. Each has a plain fix.
First, why it happens at all
ChatGPT isn't being stubborn. It works out what you most likely want from everything it can see: the current message, the earlier chat, your custom instructions, any Project instructions. Your rule is one input among many. If the rule is fuzzy, contradicted, or sitting far back in a long chat, something else can win.
So every fix below is the same idea in a different form: make the rule the clearest thing ChatGPT can see.
1. The rule is too vague
"Make it more professional" gives ChatGPT nothing to aim at. It rewrites things you liked and leaves the real problem.
Fix: turn adjectives into checks. Instead of "short," say "under 100 words." Instead of "professional," say "no exclamation marks, no 'just checking in.'" A rule it can check against is a rule it tends to follow. (Still count the words yourself. Limits are close, not exact.)
2. Two rules clash
"Keep it under 50 words" and "include every detail from the notes" can't both win. Neither can "warm and chatty" and "formal legal tone." When rules collide, ChatGPT picks one, and it may not pick the one you care about.
Fix: read your prompt and your custom instructions together, out loud if you have to. When two rules pull in opposite directions, decide which one matters for this task and delete the other.
3. The rule got buried in a long chat
You set "no emoji" at the top of a chat. Forty messages later, emoji start appearing. Everything in a chat competes for attention, and the older parts weigh less as the chat grows.
Fix: restate the rule in your latest message, or start a new chat with the rules at the top. My test: if I've corrected the same thing twice and it comes back a third time, I stop arguing and start fresh.
4. The rule lives in the wrong place
ChatGPT has several homes for standing rules, and they don't all apply everywhere:
- Custom instructions apply to every chat you start.
- Project instructions apply only inside that Project, and they win when they disagree with your custom instructions.
- Temporary Chat doesn't use memory, so a rule saved only as a memory won't be there.
- Memory is the weakest home for a rule. ChatGPT decides what to keep, and it can miss or misapply it.
One more catch: editing a Project's instructions doesn't rewrite old chats. Start a new chat inside the Project to test the change.
Fix: put rules for a repeated job (client emails, proposals, weekly reports) in a Project. Put personal preferences in custom instructions. Don't rely on memory for anything that would embarrass you if it were skipped.
5. The rule only says what not to do
This is the one I see most. "Don't invent prices" is a good rule. But if the request arrives with no prices, ChatGPT is stuck. It obeys, and you get a draft with a blank where the number should be, or it fills the gap with something vague.
Fix: pair every "don't" with what to do instead. For the agency above, the line becomes:
Don't quote a price unless I give it to you.
If a price, deadline or client name is missing, ask me for it first,
and don't write the draft until you have it.
That one extra sentence changes the behavior completely. Now a request with a missing price gets a question back, not a guess.
6. The rule needs a fact ChatGPT doesn't have
"Use our standard terms" can't work if the terms aren't in the chat. "Match our tone" can't work if ChatGPT has never seen your tone.
Fix: give it the facts the rule depends on. Paste the rate card into the Project files. Paste two or three real emails as examples of tone. A rule plus its facts holds. A rule alone is a wish.
A related trap: it copies your examples too well
Examples are strong signals. Paste three past client emails to teach your style, and ChatGPT may also borrow a name, a date or a promise from one of them. I add a guard line whenever I use examples:
Copy my style, not the facts. Don't reuse names, reasons or promises
from the examples. Only use what's in this request.
How I test a set of rules
Writing rules is half the job. The other half is trying to break them. My routine:
- Put the rules in a Project.
- Open a new chat in that Project.
- Send a one-line request that is deliberately missing something (no price, no client name).
- Read what comes back. If it guessed, the rule needs a "do this instead" line. Add one line, start a new chat, and test again.
Ten minutes of this beats a month of correcting drafts by hand.
The checklist
- Be specific. Numbers and names beat adjectives.
- One rule per line, so you can find and fix the one that slips.
- Say what to do, not only what to avoid.
- Remove clashes between the prompt and your custom instructions.
- Keep standing rules in a Project, not memory, and test them in a new chat.
- Read the result before it goes to a client. Good rules still slip now and then.
Try it step by step
The full write-up, with all six causes and where each rule belongs, is here: Why ChatGPT Ignores Your Instructions (and How to Fix It).
If you want to practice the "test, find the gap, add one line" routine, the free lesson Turn a Great Prompt into Reusable Instructions walks through it in a practice copy of ChatGPT. You sort a weekly prompt into rules and blanks, put the rules in a Project, and test them with a request that's missing a fact. It's free, about 7 minutes, and nothing touches your real account.
What a rule that holds saves you: the rewrite, the walk-back email and the awkward call. That's the business case for spending ten minutes on it.
