While the model is working

1. 01Output starts appearing in under two seconds, or the interim state explains what is happening.A spinner is not an explanation. “Reading your three attached files” is.
2. 02For multi-step work, the user can see the plan or steps before the result lands.A working transcript, not a progress bar. Steps a human can audit.
3. 03The user can stop generation at any point without losing what has already streamed.
4. 04The user can redirect mid-run — correct course without starting over from a blank prompt.Interrupt-and-steer is the difference between a colleague and a vending machine.
5. 05The effort the system shows matches the stakes of the task.Ten seconds of visible reasoning to retitle a doc is theater; instant confident output on a legal question is malpractice.
6. 06Long or structured output lands in a stable artifact, not an ever-scrolling chat bubble.
7. 07Anything destructive or externally visible pauses for explicit approval before executing.Sending, deleting, publishing, and paying are approval-gate territory. No exceptions for “power users.”
8. 08The user can tell, at a glance, whether the system is still working, done, or stuck.
9. 09We have watched the slowest realistic case — biggest input, worst latency — and the experience holds.Demo inputs lie. Test with the 40-page document, not the 2-paragraph one.
10. 10Keyboard users can invoke, interrupt, and accept without touching the mouse.

Discussion prompts

• At what exact moment does the human stop being the author and start being the reviewer — and did we choose that moment, or did latency choose it for us?
• If the user walks away for two minutes mid-generation, what do they come back to? Does anyone here actually know?
• Which step in the transcript would a user most want to challenge, and can they?
• Is our streaming a legibility tool or a perceived-performance trick? What would we change if we admitted which one?
• What is the most expensive action this feature can take, and how many clicks stand between the model and that action?
• Where are we showing work nobody asked to see, and what is it costing in time-to-answer?