Designing Refine, an AI feature that helps people describe the room they want
How 1.2M generations decided where the help should live, and why the answer was inside the prompt box.
Company
Renovate AI
Platform:
Mobile and web
Status:
Final prototype. Usability testing scoped, not yet run
Timeline
One intensive design cycle
Preview
01
TL;DR
Renovate AI lets people re-imagine a room from a photo and a line of text. The prompt box was blank, and a blank box assumes people can write like designers. Most of them can’t, so they typed something vague, got a weak render, and left.
Before committing to a direction I pulled 1.2M generations. Then I mapped three places the help could live and built the winner into the prompt box itself: two actions that turn what someone typed into a structured prompt, without ever taking away their ability to write it themselves.
02
Context
Renovate AI is an AI renovation and interior-design app, used by over a million people across mobile and web. It offers:
AI redesigns of your room
Photo-real renders of your space
A guided style finder
An AI design consultation and cost estimation feature
Saved, shareable projects
4.68
★★★★★ 10,830 ratings on the App Store
Web
The web app, where most long descriptive prompts get written.
Mobile
Mobile, where most targeted edits happen.
03
Before the redesign
People know the room they want. They can’t describe it.
A blank prompt box assumes users can write like designers. Vague prompts produce weak renders, so people regenerate, get frustrated, and leave.
The box people were actually facing: no examples, no vocabulary, and a keyboard waiting.
You can read the same gap in the App Store reviews. People describe asking for something and getting back something adjacent to it:
I’m sorry, but I didn’t like it ★☆☆☆☆
I tried to ask the AI to make my room clean, but they didn’t clean my whole room. All they did is just clean my closet, which made me disappointed…
Find it lack luster ★★☆☆☆
Taking one photo does not allow to encompass total space. Suggestions are often not utilizing space well. , even using halls that can’t be used.
Real reviews of Renovate AI on the App Store. They are about the app as a whole, not about Refine, which had not shipped when they were written.
04
I analysed 1.2M generations to frame the problem
Before committing to a direction I pulled the generation data. Three things stood out, and the first one reframed the whole project:
Most generations got no feedback at all. Roughly 810,000 of them, against about 215,000 thumbs down and 165,000 thumbs up. The silence is the signal.
Drop-off stayed high after every outcome. Around two thirds after a thumbs down, and still close to 60% after a thumbs up. Even a good render wasn’t holding people.
The working vocabulary was tiny. Prompts clustered on a handful of verbs, add, change, make, remove, paint, and the most repeated prompts were blunt: “remove”, “add landscaping”, “change color”.
Users don’t tell you when it fails. They just leave. That is the finding the rest of the design had to answer to.
Outcomes
Most generations end in silence, and drop-off stays high even after a thumbs up.
Most repeated prompts
The same blunt phrases, over and over: the vocabulary gap in one table.
05
Two users, one prompt box
The blank box fails two different people.
The new user doesn’t know what the system can do. The active user knows, but lacks the designer’s vocabulary to describe what they want.
The research calls this the domain-knowledge gap: you can’t ask a good question in a language you don’t speak.
06
Two problems that followed
Framed that way, the cost showed up in two places:
Every wasted generation costs money twice. Real compute cost, plus the user’s belief that the product works. The first render is the moment they decide.
New users don’t know what the system can do. The empty box gives them nothing to react to and nothing to learn from.
07
Objectives
New users need to learn what’s possible. Active users need to get better results. One surface had to do both.
08
Approach
Four rules I held the design to:
Treat prompting as the product’s job, not the user’s.
Put suggestions where attention already is: at the input field.
Specific beats vague. “Mid-century living room, warm wood” outperforms “get ideas”.
Never force the suggestions. Typing your own prompt always works.
09
Where should help live
I mapped three places the help could live and built each one far enough to judge it:
A style quiz before upload
A fix-it step after the first render
Help inside the prompt box
A quiz before upload delays the thing people came to do. A fix-it step only arrives once the render has already disappointed them. Only the third option reaches both users at the moment they are actually stuck, which is the moment they are staring at an empty box.
All three built far enough to judge. Only the third one meets people at the moment they are stuck.
10
The third option won: help inside the prompt box
Two actions sit with the input. Help me describe, for the person starting from nothing, and Improve description, for the person who knows what they want but not how to say it.
Tap either and what you typed is rewritten as a structured prompt: “Modern kitchen cabinets” becomes “Modern kitchen cabinets with white cabinets paired with matte black hardware.” You pick the version you meant. Typing your own prompt still works exactly as it did.
The same two actions ship on mobile and on web, sitting against the input in both places rather than behind a menu.
Mobile and web get the same two actions, and a tap rewrites the prompt into something specific.
11
Where it stands
Final prototype. Usability testing is scoped but not yet run, so what I have is a design grounded in the generation data rather than a result.
The thing I’d watch first is whether the rewritten prompt still sounds like the user’s own intent. The data told me people leave silently; it can’t tell me whether they’ll accept a sentence the system wrote for them.
This case study covers a feature that has not yet shipped or been tested with users. The framing came from 1.2M generations; whether the solution works is what the usability test is for.