You knew exactly what you wanted. It came back as something else.

I've been building with AI for over a year, and for most of it I had a habit I didn't notice: when a session gave me back the wrong thing, I'd sit there rewording the prompt until it finally understood me. Sometimes that took ten minutes. Sometimes it ate an afternoon.
The moment it clicked for me happened while watching our product lead, who runs product and doesn't write code. He described an onboarding screen out loud for about ninety seconds. Clean, one input, a soft nudge telling a new user what to do next. He knew exactly the feeling he wanted. He handed that description to an AI session, and what came back was a login form with four fields, a terms checkbox, and a modal that fired before the page settled. Nothing wrong with the code. It just wasn't the thing in his head.
The threads I read call this vision drift. Not "I can't code." It's sharper: I know exactly what I want, and it keeps coming back as something else.
Here's what I got wrong for a long time. I treated this as a communication problem, the kind you fix by talking better. You can't always talk better. The gap isn't syntax. It's the dozens of small decisions sitting between a description and a working screen: field count, validation timing, whether the nudge is a tooltip or inline text. On a team with a technical lead, someone with judgment stands in that gap and absorbs those calls invisibly. A solo non-technical founder doesn't have that person. The AI makes every one of those calls confidently, doesn't tell you it made them, and hands you something that looks decided when it was really just defaults you never got to weigh in on.
So we stopped trying to write the perfect prompt and built something else instead. The tool is DevKanban, our AI-native dev workflow, and its core unit is a card: an isolated workstream with its own git branch, its own cloud environment, its own live preview URL, and a quality module called Firebat that checks output against what "good" means for the task. Instead of handing his ninety-second description to one session and praying, our product lead pointed three cards at the same brief. Same words, three independent interpretations, each on its own preview URL.
Card A did the login-form-heavy thing that had frustrated him. Card B stripped it to one input, almost too bare. Card C sat in the middle with a progressive nudge. He clicked through all three real screens, and the choice made itself in under a minute: C, with B's restraint on the copy and one interaction detail from A. He never specified the technical decisions. Each card made them differently, and seeing three sets of made-decisions side by side turned "how do I specify this" into "which of these do I want," which is a question he could answer instantly.
The honest part: this took a month, not a weekend, and the reason is a failure we hit while dogfooding. With several cards active, a fix we'd just merged bled into a still-running card's preview, because the environment wasn't fully sealed after the merge. For a moment one card was showing behavior that belonged to another. That's exactly the failure a real user would hit constantly, and it would have made the whole comparison a lie. You can't compare three interpretations if changing one quietly changes another. We only caught it because we were depending on the tool for real work, not testing it in the abstract. Isolation is the boring foundation under the nice-sounding feature, and Firebat exists for the same reason: harness engineering, building structure around model calls so output is reliable, is standard practice now, and this is just our implementation of it.
Where things stand: we've been living inside DevKanban for the past month, building DevKanban with it, and the parallel-cards approach is what's carrying most of that work.
So here's my question for other builders, especially the non-technical ones: when your idea comes back wrong, do you keep rewording the prompt until the AI finally gets it, or do you wish you could just see two or three honest attempts and point? And if you've found a way to close that translation gap without a technical co-founder, I genuinely want to know what it is.

We're a small team in Seoul building AI-native tools, and we run our own work on them before we ship.