The habit that made Claude Code actually useful for me: build the throwaway version first, on purpose

For a while I fought Claude when it built something different from what I had in my head. I'd correct it, steer it back, spend an hour getting the "right" architecture out of it before I'd even seen the feature work. I do the opposite now. First pass, I let it build the fast, ugly, probably-wrong version. Hardcoded values, one giant file, no tests. I'm not trying to keep any of it. I just want to watch the thing run so I can tell whether I actually asked for the right feature. Half the time, seeing it run tells me I didn't. The feature I described isn't the feature I wanted. I only find that out by using it, and I can use a throwaway in ten minutes. Then I delete it. All of it. And I describe the second version with everything I learned watching the first one fall over. That version usually lands close to right, because now I'm specifying from experience instead of imagination. The thing I had to get over was the waste. It feels dumb to have Claude write two hundred lines you're going to throw away. But those two hundred lines were never the expensive part. My wrong idea about what to build was the expensive part, and the throwaway is the cheapest way I've found to kill it early. The old way, I'd polish the wrong feature for most of a day before noticing. Now it falls apart in the first ten minutes, which is exactly what I built it to do. Does anyone else build a deliberate throwaway pass first, or do you trust the first real attempt?