AI writes fast. It doesn't know what's good.
Key takeaway: When the first draft misses, generate a few independent attempts and choose the one that best fits your codebase.
A model’s first answer is one draw from many plausible ones. Instead of patching a draft that went wrong, run a few agents on the same spec, compare them, and pick the one closest to right. Picking still takes taste.
Run the same prompt twice and you get two implementations, with different names, structure, and sometimes a different approach. When the first one misses badly, the habit is to fix it in place, one correction at a time. Every fix gets stacked on the wrong idea, and the history of what didn’t work crowds the context toward the dumb zone. Drawing a few fresh attempts is often cheaper.
Give each agent the same spec and its own git worktree, so they can’t step on each other or see each other’s work. The spec still matters. Five agents on a vague prompt just give you five vague versions. Keep the count small, too. Another version costs tokens, and comparing it costs your attention, which is the expensive part.
Then read them side by side. Where they all agree, the problem probably had one obvious answer. Where they diverge, there’s a decision. One adds a table and another reuses a column. You might not have known that was a choice until you saw both.
All of them might pass the tests. Green tells you a version is a candidate, not which one to keep. “Closest to right” means the one that follows the codebase’s patterns, has the smallest diff, uses names that match how the team talks about the domain, and skips the dependency nobody needed. An agent can’t judge that, because it doesn’t know what good looks like in your codebase.
After picking, I tweaked the code by hand. When the gap is small, editing is faster than describing the edit. If you keep making the same tweak feature after feature, put it in a SKILL.md file so the next round starts closer.