Every AI creation tool eventually ships a chat box, and most of them quietly fail. Not because chat is the wrong interface, but because chat smuggles in an assumption about what the user is doing: that each message is a step along a path toward one known destination. Turn by turn, refinement by refinement, converging on the answer.
That’s the right model for editing. It’s the wrong model for the phase of creative work that comes before editing, and conflating the two is one of the most common design mistakes in AI products right now.
Working on Mattoboard (an AI-powered interior design platform), we kept running into a mode of use that didn’t fit either of our existing UIs. Watch a designer (or writer, or anyone doing creative work) and you’ll see three distinct jobs:
Creation: “I have an idea, give me a starting point.” One shot, fresh canvas, no history to reconcile against.
Tweaks: “Change this wallpaper to linen.” Surgical edits on an established direction. Everything else stays untouched. This is where a refinement chain usually fits.
Exploration: “I know the direction, not the destination.” The user is holding a tension they haven’t resolved: traditional but modern, warm but restrained, formal but social. They don’t want the answer. They want to see how different it could be, a few genuinely distinct takes, so they can figure out what meets the expressed desires.
That third job is where chat interfaces break. The user types “make it warmer” and the tool dutifully patches the previous image, walking one step down one path. But the user wasn’t asking to walk. They were asking to sample. They wanted four warmths, side by side, so they could locate their own taste by comparison. The tool converged; the user needed to triangulate.
The navigation term is literal here. Triangulation is how you locate a point you can’t see directly: take bearings from multiple known landmarks, and the intersection tells you where it is. Your own taste is that unseen point. Every sample you react to is a landmark (“not that, too formal” is a bearing), and enough spread-out bearings locate what you want. Convergence, by contrast, is dead reckoning: walking one step at a time toward a destination you’ve never seen.
An earlier version of Mattoboard’s generation feature had explored chat and it hadn’t landed. The postmortem instinct was “chat was wrong.” The real diagnosis was subtler: the UI implied a refinement chain (converging) while users were trying to sample a space (triangulating). Same input box, opposite mental models, and the mismatch made every response feel slightly off.
A convergent chat UI is seductive to build. It maps cleanly onto conversation history, it’s cheap (one generation per turn), and it’s a “sexy” demo: it looks like a clean way to hand over the answer upfront, without the user doing any work. In the current zeitgeist, it feels familiar and timely. But it fails the actual user in two ways.
First, it never shows you options at all. Every turn produces exactly one output built on one interpretation. If the model read “warmer” as color temperature and you meant emotional warmth, you’re now three turns deep in the wrong branch, and the interface offers no way back except starting over.
Second, and worse, rejecting an output replaces it instead of keeping it. Regenerate usually rerolls the last turn, and the version you just rejected vanishes from the working surface. The history may still sit in the chat log, but you can’t hold it next to the new attempt and compare. Physical creative work operates the opposite way. An interior designer pulling samples doesn’t clear the table between options; the rejected marble stays next to the candidate oak, because rejected options are information. They’re the landmarks you triangulate from. Digital tools took a natively cumulative workflow and made it one-in, one-out.
A few principles that fall out of taking the sampling model seriously:
Anchor to intent, not to the last output. In a convergence UI, each turn’s semantic anchor is the previous image. In a triangulation UI, the anchor should be the brief (the user’s stated tension) and each generation is an independent sample against it. “Warmer” shouldn’t mean “patch the last image warmer”; it should mean “resample the space along the warmth axis.”
Make the axes explicit when you can. Sometimes users know exactly which dimension they’re exploring (style, colorway, layout) and sometimes they only have prose (“torn between formal and social”). Parse the prose into named axes, show the user what you heard, and let them pin or correct. The parse itself becomes a moment of clarification that helps the user understand their own ambiguity.
Keep rejected samples within reach. The working view should stay focused on the options you’re actively comparing, but going back should always be one gesture away. A bad sample tells you where the edge of your taste is; throwing it out throws out that information.
Keep trying things out separate from the real thing. Nothing you generate should land on the actual board until you say so. That keeps experimenting cheap (nothing counts yet) and the board honest (everything on it was chosen).
This isn’t just an interior design problem. It’s the same reason “regenerate” buttons feel like gambling, why prompt-and-pray image tools exhaust people, and why the best AI writing workflows produce three drafts to react against rather than one draft to fix. Human taste is comparative. We mostly can’t describe what we want, but we’re extremely good at ranking options once we see them. Comparison gets a bad reputation (the thief of joy, and so on), but in creative work it’s the opposite: comparison is where the joy is. Reacting to options is the fun part; describing what you want from nothing is a chore.
AI tools that only converge ask you to explain what you want before you’ve seen a single option. AI tools that triangulate let you point at it once it’s in front of you. People are bad at the blank page and great at reacting. Reacting is almost always the better bet, and it changes what you build: parallel generation for determining direction, persistent context over resets, sample surfaces over refinement chains. Building towards natural human tendencies creates a more delightful experience.