Not every off result is worth correcting in place. Sometimes 3B handles it, sometimes the faster fix is to start over. This helps you tell which is which.
Most of the time, iterating is the right call: the build’s close, and one specific piece of feedback closes the gap. But some problems aren’t yours to fix at all, and every so often a conversation wanders far enough that you’d actually be better off starting fresh than trying to rescue the thread. The skill is spotting which of the three you’re in before you pour more time into the wrong one.
Sometimes the fix isn’t yours to make
Before you dive in, it’s worth checking whether the problem is even yours. 3B can catch and sort out a fair few things on its own. When a step fails or produces something malformed, it’ll often repair it and carry on rather than handing it back to you. So give it a chance to clear the obvious breakages itself before you start correcting. The fixes worth your hands are the ones about intent, where the build runs fine but isn’t doing what you meant. That’s the gap 3B can’t close alone, because you’re the only one who knows what you were going for.
Signs you should keep iterating
Stay in the conversation and lean on feedback when:
- The result’s close and the problems are specific things you can name.
- Each round of feedback visibly moves it nearer to what you want.
- The bones are right and you’re refining details, wording, or edge cases.
- You can still clearly explain what you want, and 3B is following you.
Signs you should start fresh
Think about resetting with a new prompt when:
- You’ve given the same correction more than twice and it keeps missing.
- Fixing one thing reliably breaks another, and you’re going in circles.
- The build has drifted so far from the original idea that you’re not quite sure what it’s doing anymore.
- You’ve realized the outcome you described at the start wasn’t the one you actually wanted.
Why starting fresh often wins
Restarting feels like tossing out progress, so people resist it and keep patching. But a long conversation drags all its earlier assumptions along with it, including the ones that sent it sideways in the first place, and no amount of feedback fully clears them out. A new prompt, written with everything the failed attempt just taught you, often gets you a cleaner result faster than ten more corrections would. And that first attempt wasn’t wasted, it’s the reason the new prompt is sharper.
When in doubt, notice which of three things is happening. If a step just broke, give 3B room to fix it. If the build’s close and you can name what’s off, keep iterating. If you’re going in circles, take the hint and start fresh.
Next up: Reading results critically, on checking whether a build actually does the job, not just whether it looks the part.