Which tool to choose
The tool is the personal instructions section, because a prohibition has to apply in all chats, not just one. Write the limits in the positive where you can ("always answer concisely") and in the negative where needed ("never open with a recap of the question"). Automatic memory isn't the right place for boundaries: they're too important to leave to chance. A limit that matters you write yourself, in plain text, where you know the assistant always reads it.
How to do it
- Think of the AI behaviors that annoy you and that you correct often.
- Turn them into sharp rules in the instructions: "don't do X", "always avoid Y".
- Be concrete: better "don't use emoji" than "be professional", which leaves too much room for interpretation.
- Check with a few normal requests that the limits are respected, and add the ones you realize you forgot.
A concrete example
Federico couldn't stand that the assistant, before answering, would repeat his question and close with a recap. He wrote two prohibitions in the instructions: don't repeat my question before answering, don't add final summaries. Since then the answers go straight to the point and end when the content ends. It only took putting in writing, once, two annoyances he used to correct by hand in every single chat.
When it DOESN'T work (and how to fix it)
If the AI breaks a limit now and then
It happens, especially in long conversations, where it tends to return to its habits. When it does, call it out on the spot ("I'd asked you not to do that") and, if the limit is important, put it back at the top of the chat. A repeated boundary holds better than one said only once.
If a prohibition is too vague
"Don't be boring" isn't a limit the AI can respect, because it doesn't know what you mean. Make it concrete: "don't make lists longer than five points", "don't use big words". The more precise the prohibition, the more applicable it is.
If the limits become too many
An endless list of prohibitions confuses and contradicts itself. Keep only the boundaries that truly matter, the ones tied to real, recurring annoyances. A few sharp limits carry more weight than twenty rules that no one, not even the AI, can hold together.
A tip from someone who really uses it
Build the list of prohibitions over time, starting from the real annoyances. Don't sit at your desk imagining everything the AI could do wrong: the list would come out long and useless. Every time an answer irritates you and you find yourself correcting it, stop and add that "don't do" to the instructions. Within a few weeks you have a handful of boundaries that reflect exactly your irritations, and the assistant stops trampling on them. The best limits are born of experience, not imagination.
Frequently asked questions
Better to write limits in the positive or the negative?
Where you can, in the positive ("always answer briefly"): it's easier to follow. But some annoyances can only be expressed as a prohibition ("never use this tone"), and it's right to write them that way. What counts is that they be concrete, not the form.
Do limits apply even with memory off?
Personal instructions are a different thing from automatic memory: as a rule they stay active regardless. If in doubt, put the most important boundaries back at the start of the chat, so you're sure they apply.
Will the AI always respect my prohibitions?
Almost always, if they're clear, but it's not infallible: in long chats it can slip. Keep an eye on the limits you care about most and call it back when needed.
Does the AI understand on its own what's best not to do with me?
No, and it's the expectation that generates the most daily annoyance. The assistant has no way of knowing what irritates you: it starts from an average behavior, meant to suit everyone, which for you may be exactly the wrong one. Implicit limits don't exist for the AI: only the ones you write do. Until you put your boundaries down in black and white, you'll keep correcting the same behaviors in every chat, wondering why it doesn't learn: it doesn't learn because you haven't told it.