Interrupting an agent mid-generation — hitting Escape and issuing a correction — the moment it's visibly heading in the wrong direction, instead of letting it finish an unwanted output and fixing it afterward.
Letting a mis-directed agent run to completion wastes the time spent generating the wrong thing and can compound if later output builds on the wrong earlier output. Interrupting early keeps the correction cheap. This is the live-steering counterpart to One-Shotting as a Workflow Goal: one-shotting is the goal for well-specified tasks, interrupt-and-steer is the fallback when the spec turns out incomplete and the agent's live trajectory reveals it.
Watch agent output as it streams rather than only reviewing the finished result; the moment it looks wrong, interrupt immediately rather than waiting for the full generation.
"Escape all caps no."
Ryan Nystrom (Notion) — personal habit described in the interview.