Steer It Mid-Run, Don't Restart
You never have to sit and watch. Type a correction while the agent works and it folds your note into the run without starting over; Esc stops it dead when you need the brakes.
Why it works
New users treat a running agent like a launched rocket — untouchable until it lands, then rebuild from zero if the trajectory was off. It's actually more like a contractor on a call: talk to it while it works. Mid-run corrections preserve everything already done and cost one sentence; stop-and-restart costs the whole run.
How to do it
- Give a task and watch the live transcript as it works.
- See it drifting — wrong color, wrong file, wrong assumption? Type the correction immediately.
- Send it. The note joins the run; the agent folds it in without starting over.
- Stack corrections as you spot things — each lands mid-flight.
- For a real emergency, press Esc or click Stop: it halts instantly, work-so-far preserved.
- After a stop, continue with an instruction, not a restart: 'keep the layout, redo only the header.'
Copy this
prompt
Correction while you work: use the blue from the logo (#2200FF), not the default blue, and keep all headings in the same font as the body.
What to watch for
- Waiting for the run to finish before mentioning a visible mistake buys you nothing — the mistake compounds into everything built after it.
- Don't spam stops for cosmetic issues; a mid-run note fixes those. Stop is for 'wrong direction entirely'.
Sources
Original: conon.ai