Standing instructions: a profile that actually works

Every major assistant has a field, often called "custom instructions" or "preferences", whose contents are added to every one of your conversations. Filled in well, it saves you repeating the same details a hundred times. Filled in badly, it quietly degrades all your answers.

What deserves to go in it

• Who you are, in two lines, only what changes the answers: your job, your level in the fields where you ask questions, your country (without which questions about law or tax often come back based on foreign rules).
• How you want to be answered: length, tone, structure, language, units.
• What you no longer want to see: introductions that restate your question, bullet points everywhere, pleasantries.

What has no place there

Sensitive information (health, address, bank details): it would go into every conversation, including those that have no need of it. And empty instructions like "be precise" or "don't get it wrong": they change nothing, the model is already trying.

The trap: an instruction applies everywhere

"Always answer in three sentences" is perfect for your quick questions, and disastrous the day you ask for a full report. Write conditional instructions: "For a factual question, answer briefly. When I ask for a text to send, give the text directly, with no commentary around it."

Short, tested, revised

This text is re-read with every message, and a forty-line block dilutes what matters. Aim for about ten lines. Then test it: make three of your usual requests before and after, and compare. An instruction that changes nothing in practice gets deleted.

Key points

• Permanent instructions are added to each of your conversations.
• Put in what changes the answers: useful context, format, what you no longer want.
• Nothing sensitive, nothing empty, and conditional instructions rather than absolute ones.
• About ten lines, tested on your real requests.