Principles

Eight rules that settle most design questions. If two disagree, the earlier one wins.

1. Principle 1

   Make AI authorship visible

   In a chat, the AI's replies are open text and the person's are bubbles, and screen readers hear "Assistant" before each reply; the small lime mark names the AI once at the top of a thread. Anywhere else, AI text carries the AI badge. People should never have to guess what a person wrote and what the AI wrote.

2. Principle 2

   Plan for waiting

   Show a typing sign before the first word. Show text as it is written, with a caret. Hold space with placeholders. Always offer Stop.

3. Principle 3

   Make the assistant's work easy to follow

   The assistant sums up its work in one line. Details are one click away. Anything that matters waits for approval.

4. Principle 4

   Help people know how much to trust

   Name sources. Show confidence with words and a shape together. Put a plain notice near the message box.

5. Principle 5

   Fail gracefully

   Errors say what happened and offer a retry. Refusals stay neutral and suggest another way. Usage limits say when they reset.

6. Principle 6

   Keep the person in control

   Deny is as easy as Approve. People can stop, try again, flip through versions and give feedback.

7. Principle 7

   One job per surface

   If the tray already shows who needs you, the side list does not repeat it. Remove before you add.

8. Principle 8

   Accessible by default

   Every icon-only button has a name. New replies are announced. You can always see where you are. Motion follows the device setting.