AI command center for homeowners. It defines the scope, compares contractor quotes apples to apples, and pauses for your approval before anything happens.
Professional
May – Jun 2026
Lead Product Designer (Me)
2 Product Designers
7 Engineers
3 PMs
React + Tailwind
Claude Code
Problem
It is one of the most stressful purchases a homeowner makes, and eight interviews showed why. Trust starts social: people ask a neighbor or a friend first, because they assume online ratings can be gamed. When word of mouth runs out, they fall back on marketplaces they trust even less. Then the grind begins. They repeat the same problem to contractor after contractor, book visit after visit, and wait on quotes that come back impossible to line up against each other. Three frustrations surfaced in almost every interview.
No way to know who to trust
"Finding the right person is the hardest part."
No benchmark for a fair price
"You can't really know if the price is fair."
Comparing quotes is exhausting
"I had to make so many phone calls just to compare prices."
Beyond our eight interviews
Leaf Home / Morning Consult, 2025. HomeServe, 2025.
Insight
Every homeowner unsure about price was actually unsure about scope. When contractors propose different fixes for the same issue, their quotes describe different work, so there is nothing to compare. One homeowner was quoted $250 to $7,000 for the same siding repair, and the telling part is that he had already worked out the answer: hire someone to diagnose and define the scope, then send that one scope out for comparable bids. He had described Homewise before we built it. Fix the scope, and the price becomes legible.
Same siding repair · 5 contractors quoted
Solution
Homewise acts on the reframe: scope first, price second. It turns a plain language problem into a defined scope of work, sends that one scope to every vetted contractor, and brings the bids back on identical terms, pausing for your approval before anything moves.
Two things make hiring hard, and neither is the automation. Homeowners rarely know how to describe a repair, and they have no reason to trust the numbers that come back. A chatbot would leave them guessing at both, so I pushed for a structured workflow instead.
| Stage | What the workflow does |
|---|---|
| 1. Scope | A short intake conversation and a photo become a structured, editable scope of work, so you never have to know the right words. |
| 2. Match | It finds and vets contractors, surfacing verified pros and the ones other Homewisers recommend. |
| 3. Compare | One scope means comparable bids. Outliers and gaps get flagged, and you approve the right contractor. |
| 4. Track | The job stays in one place from booking to done, so it never slips back into scattered calls and texts. |
Process
Engineering used to be the slow part, so design got front-loaded and the build came last. Now the team can stand up a working version faster than I can finish a static mock, so I flipped the order and build first. Each loop starts from the PM's PRD, our shared intent: the problem, the users, and the metrics to move. The prototype exposes gaps the document cannot, which feed back into the PRD. Homewise especially needed this, since one change ripples through scope, matching, quotes, and tracking, so I validated it live. Less time on mockups, more time on the calls only a person can make. I took this from Jenny Wen, Head of Design for Claude, who argues the design process is dead.
Final Design
You describe the problem in your own words. The agent turns it into a structured, editable scope of work, the most important artifact in the whole product.
The same scope goes out to every contractor, so you stop repeating yourself on the phone and everyone quotes the exact same job.
Homewise surfaces a verified, fairly priced match with its reasoning: a match confidence, verification on license, insurance, and real permit history, and how many Homewisers recommend them. Social proof, the way a neighbor refers someone, not an anonymous star rating.
Bids line up side by side with a plain language summary and outlier flags, the comparison homeowners reach for ChatGPT to fake today. The agent recommends, but nothing happens until you approve. A deliberate, human moment, never an autopilot default.
Once a contractor is booked, the whole job lives on one dashboard. It tracks itself from scheduled to closed out, so it never slips back into scattered calls and texts.
Homewise is live and interactive. The whole flow runs end to end, from the first plain language request to the moment you approve the hire.
Design System
Homewise runs on Hearth, its own design system. It reads like a calm, premium homeowner tool, closer to Notion or Linear than a SaaS dashboard, with a humanist warmth most AI brands skip.
The palette is read by role, not hue: sage for trust, ember for caution, sky for info, over a warm cream canvas and near black ink. Type is one typeface, Geist, on an editorial weight ladder capped at 52px, because larger sans starts to shout. Shadows are banned, so depth comes from a five tier surface ladder and hairlines instead. And it is living code, not a static spec: every new token ships behind a drift linter.
Collaboration
Working closely with the PM and engineers, we made an unusual call: skip Figma and design in the codebase from day one, a real React prototype on Hearth. We handed the dev team the repo itself, and they built the real backend straight onto it. Skipping Figma did not close the door on it. If the team ever needs a traditional handoff, I can push the prototype from Claude Code into Figma through its MCP server, turning the running interface into editable frames. The code stays the source of truth, and Figma is a step we reach for only when it earns its place.
A full walkthrough of the shipped build, start to finish.
Reflection
The hard part was never the automation. It was trust. People only let the agent run when they could see the scope it built, the contractors it verified, and knew the final yes was still theirs. Trust was a data problem as much as an AI one, credible only because the matching ran on real license and permit records, not the model's best guess. Building a front end prototype in Claude Code for the first time reframed the deliverable too. The design wasn't a picture of the product. It was the first version of it.
Powered by Claude Haiku 4.5