Delete the work and start again

When an AI session gets me 80% of the way there, I delete it and start over. Not sometimes. As a rule.

The instinct to keep going is strong. You’re close. A few more exchanges and you’ll get it across the line. But sessions that start losing shape don’t recover. I’ve watched it enough times to stop pretending otherwise. Once the context drifts, every follow-up prompt is fighting the session’s momentum instead of steering it. You’re negotiating with the wreckage of your earlier asks.

So I stay a bit longer, but not to fix it. I stay to study the shape: where did it go right, where did it buckle, what was I actually asking for when I thought I was asking for something else? The session won’t produce the answer. It never does once it’s lost. But it will show me what I was reaching for, and that’s all I need to frame the next attempt. I’m reading the chart before I close it.

Then I delete the work and start again.

This only makes sense once you let go of a belief left over from when producing code took real time and real resources. Back then, throwing away a day’s work meant throwing away a day. Now the cost of a session is minutes. The sunk cost instinct still fires, but there’s nothing behind it. The sunk cost isn’t compute. It’s the clarity you sacrifice by nursing a mediocre session instead of applying what you learned to a clean start.

What survives each deletion is the ask. Every failed session sharpens how I frame the next one. The code is disposable. The clarity isn’t. Over weeks and months of this, the asks compound. You stop getting 80% of the way there and start landing clean on the first attempt.

The only edge left is getting better at how you ask. Deletion is how I practice.