Which tool to choose
The main assistants all have them, with slightly different names and forms. On ChatGPT they're two distinct boxes, one for the context about you and one for the way of answering. Claude and Gemini offer equivalents in the personalization settings. Choose the one you already use: the concept is the same, only where you find the setting changes.
How it's done
- Open the settings and look for the custom instructions or personalization section.
- In the first part write the stable context: role, sector, who your texts are aimed at.
- In the second write how you want the answers. The operational syntax:
Answer in English, in a direct and concise way, without preambles like "great question". Use bullet points when you list several things. If you're missing information to answer well, ask me for it instead of guessing. Keep a professional but simple tone.
- Save, try it on a few requests and adjust: if something doesn't work, it's usually one rule too many or two that contradict each other.
A concrete example
A lawyer was tired of getting answers full of jargon that he then had to translate for clients. He put two things in the custom instructions: that he writes for people without legal expertise, and that he wants explanations in plain language with a concrete example for every concept. From that moment every chat, with no further indications, gives him back texts already suitable for clients. He stopped re-explaining his audience at every conversation: the rule, written once, works for him everywhere.
When it does NOT work (and how to fix it)
If the AI doesn't always respect the instructions
Almost always it's because you've put too many, or two that contradict each other ("be concise" and "explain everything in detail"). The fix is to trim: keep a few clear, compatible rules. Three sharp instructions are worth more than twenty that quarrel.
If the right behavior changes from job to job
Custom instructions are global, they apply to everything. If a specific project needs its own rules, don't upend the global ones: use a project space with its own dedicated instructions, leaving the global ones for what always holds.
If in a particular case it ignores them
On very specific requests it can deviate. The fix is simple: restate the rule at the start of that chat ("remember to answer concisely"). The targeted repetition, for the single case, sets things right.
A tip from someone who really uses it
Among all the rules you can set, one pays off most of all: "if you're missing information, ask me for it instead of making things up". Put in the custom instructions, it reduces guessed answers and fabrications upstream, because it pushes the AI to stop and ask instead of filling in the gaps. It's the single instruction that improves everyday quality the most.
Frequently asked questions
Do custom instructions apply to all chats?
Yes, they apply to every new conversation until you change them. That's exactly their purpose: to give you consistent behavior without repeating yourself. They stay active in the background, even when you don't think about it.
Are they the same thing as memory?
No. Custom instructions are rules you deliberately set; memory is material the AI gathers on its own from the conversations. One you decide, the other forms with use. They work well together, but they're two distinct things.
How many instructions can I write?
There's a space ceiling, but the real problem isn't the length: it's the coherence. A few clear rules beat a dense list. If you feel the need to write a lot of them, some probably contradict each other and should be simplified.
The more detailed and longer I write them, the better the AI behaves?
No, and it's the most common mistake. Beyond a certain point, long, dense instructions start to contain conflicting requests, which cancel each other out and leave the AI halfway. Quality doesn't come from the quantity of rules, but from their clarity and the fact that they don't step on each other's toes. Fewer and sharp works better than a lot and tangled.