Iterate and Improve: Good Software Takes Loops, Not Leaps
Claude Code's first result is a starting point. Turning it into something solid takes a loop: read what it built, ask for specific changes, and try a different approach when refining stops helping.
The first-try fallacy
You've seen the demos: someone types a prompt, AI produces a working application, everyone applauds. Then you try it yourself. The first result is close but not right. You tweak the prompt, try again, and get something different but still not right. Three attempts later, you're wondering whether you're doing something wrong. Needing several tries is how this works, however experienced you are. A demo short enough to watch has no room for the middle part.
You might expect using AI to be a single pass: describe what you want, get what you described, ship it. When it doesn't go that way, it feels like failure. Then you either accept rough output, because you assume AI should have gotten it right, or you get frustrated and give up, because 'AI doesn't work for me.'
Software gets built in passes, with or without AI. AI makes each pass faster.