Which tool to choose
Indicating what to avoid works with any AI assistant for text responses. A useful clarification: in image generators there's a dedicated field actually called "negative prompt", where you list the elements to exclude. In text chats there's no separate field: the prohibitions go written inside the normal prompt. This guide is about text; for images the principle is similar but the tool has its own dedicated space.
How to do it
Prohibitions serve to correct recurring flaws, the ones that come back no matter how you rephrase. But a blunt prohibition sometimes has the opposite effect: the AI focuses on the forbidden thing. That's why it's worth, where possible, saying what to do instead of what not to do.
- Identify the flaw that keeps coming back: too long, too formal, full of clichés.
- Write the precise prohibition, not the generic one: not "don't be boring", but "no introductions, get straight to the point".
- Where you can, turn the prohibition positive: "use sentences under twenty words" works better than "don't make long sentences".
- Keep the prohibitions few and targeted: three sharp prohibitions count more than ten fuzzy ones.
- Check the result: if the flaw remains, make the prohibition more concrete or add an example of how it should be.
The operational syntax:
Write the description of a farm stay for the website, about 100 words.
To avoid:
- inflated adjectives like "unique", "magical", "unmissable"
- exclamation marks
- opening with "imagine that..."
Instead of these: concrete and verifiable details (what you see, what you eat, how far it is from the center).
After the response, check exactly the elements you'd forbidden. If one remains, flag it: "you used 'magical', remove it". The targeted correction on a single prohibition is more effective than a new prompt from scratch.
A concrete example
Chiara has captions generated for social media and the AI always slipped in three emoji and a final "find out more!" that she hated. She adds two precise prohibitions to the prompt: no emoji, no shouted calls to action. And she turns the rest positive: "close with a real question to the reader". The captions come out clean already. The flaw she fought by hand on every post disappeared because she wrote it once as a rule.
When it does NOT work (and how to fix it)
If forbidding a thing makes it appear more
It happens: focusing attention on the forbidden element sometimes summons it. Rephrase positively. Instead of "don't use a pompous tone", write "simple tone, as you'd talk to a friend". Describing the wanted result is more effective than indicating the one to avoid, because it gives the AI a direction instead of a hole.
If you put in too many prohibitions and the response becomes rigid
A long list of "don'ts" intimidates the AI and produces stilted text. Keep the two or three prohibitions that really matter and let the others go. The minor flaws you correct afterwards, you don't prevent them all at the start.
If the prohibition is too vague to be respected
"Don't be generic" doesn't tell the AI what to do differently. Make it concrete: "every statement must have an example or a number". A prohibition the reader wouldn't know how to respect is too vague even for the AI.
A tip from someone who actually uses it
Keep a small personal list of the flaws your AI repeats: your tell-tale phrases, the tics you recognize at a glance. When you spot a new one, add it to the fixed prohibitions in the custom instructions. Over time you build a filter tailored to the flaws that bother you specifically, and you stop correcting the same ones over and over.
Frequently asked questions
Is it better to say what to do or what not to do?
When you can, what to do: the positive instruction gives a direction, the prohibition leaves a void the AI fills at random. The prohibition stays useful for the specific, recurring flaws you want to cut ("no emoji"). The two combine: a few positive rules on the result, some targeted prohibitions on the tics.
Does the image negative prompt work like the text one?
The principle is the same (exclude unwanted elements), but in image generators there's a dedicated field to list the exclusions, while in text the prohibitions go written in the normal prompt. They're two tools with the same idea behind them.
If I list enough prohibitions, do I get the perfect text?
No, and it's the misunderstanding that leads to endless lists. You can forbid infinitely without ever saying what you really want: the result is a text that avoids the errors but doesn't hit the target. The direction is given by the positive instructions; the prohibitions refine, they don't build.