Instructions are standing preferences the AI Assistant applies to every conversation.
Set them once, and the Assistant follows them until you change them without having to repeat yourself in each new chat.
Use them to tell the Assistant how you like to work: how to address you, the tone to use, or how you want responses formatted.
đ Instructions are personal to you. They apply to your conversations with the Assistant and don't affect anyone else on your team.
đ Your assistant also picks up details about you and your work as you chat, without you having to write them down (Beta). To see how instructions and memory differ, see AI Assistant: Memory.
Setting Your Instructions
Go to Settings > AI Assistant.
In the AI Assistant, open Settings in the chat input and select Personalization.
Your instructions are in the Instructions section, next to your assistantâs avatar.
Don't see this page? Additional customization options for your assistant are in Beta, so you might still see the previous screen.
If so, open Settings in the chat input and, under Personalization, select Instructions. Everything else in this article works the same.
For more ways to customize your assistant, see Customizing Your AI Assistant.
â
What to Write in Your Instructions
Good instructions capture the preferences you'd otherwise repeat in every chat. Such as:
How to address you: "Address me by my first name."
Tone: "Keep responses concise and professional."
Format: "Prefer short bullet points over long paragraphs."
Example: "Address me by my first name. Keep answers short and to the point, use bullet points, and avoid jargon."
Write your instructions in natural language, up to 10,000 characters, then click Save.
Changes apply from your next conversation, not the one you already have open.
Your instructions stay in effect across all your future conversations with the Assistant until you edit or remove them. To remove them, delete the text and click Save.
đ Click Check with AI to have the AI Assistant review your instructions, even before you save. It flags anything it can't do with its tools, connectors, and permissions, and suggests a corrected version.
