Akshay DeshpandeDiscovery
A design brief with two modes: Read for researching buildings, Write for fast reviews on a phone. It added the building profile, the missing centrepiece.
Renteroo: honest building reviews for Australian renters, designed with AI
A building-level rental review platform, and a test of what AI-assisted design looks like end to end: from PRD to a full redesign with a multi-agent design workflow.
- Role
- AI-first designer and product manager
- Type
- Personal experiment
- Tools
- Claude Code, DesignPowers skills, Figma MCP, HTML/CSS
- Status
- 8 prototype screens complete
Renters sign blind
During a recent stay in Australia, I heard several stories from family and their friends about the rental situation in cities like Sydney and Melbourne. Renters sign year-long leases with almost no reliable information about the building. There are no ratings and no reviews from past tenants. They compete with 30 to 50 applicants and decide after a 15-minute inspection of the apartment shared with everyone else.
- < 1.5%national vacancy rate since 2022
- 12 mo.leases signed after a single inspection
- 0structured building review platforms for Australian renters
Problem statement
How might we help renters make informed decisions before signing, and let them share their experience without legal risk?
What's the legal risk? More on that below.
Phase 1: From problem to first prototypes
I used the idea as a test bed for AI-assisted design. I described the problem to Claude Code, had it write a full PRD, then edited the PRD by hand where it had made wrong assumptions. Competitors like Domain, Google Maps and RateYourLandlord covered listings, places or landlords. None reviewed buildings.
The key decision
Design around the law, not against it
Before designing anything, I asked: if renters need this so badly, why doesn’t it exist? The answer was legal, and it made me rethink the platform.
In the US, a law called Section 230 protects websites from being sued over what their users post. If someone writes a false review on Yelp, the reviewer is responsible, not Yelp. Australia has no such protection. In 2021, the Australian High Court ruled in the Voller case that news outlets were legally responsible for defamatory comments strangers left on their Facebook pages, even comments they hadn’t seen. For a site where renters review landlords, that’s a serious risk: one angry, inaccurate review could make the platform liable for defamation.
So I turned the legal risk into five design principles that every screen follows. In short: host, don't author. Facilitate, don't curate.
- 01 First-person framing “How did you find the noise levels?”, not “How noisy is the building?”
- 02 Structure over free text Star ratings and tags carry the review. Free text is supplementary.
- 03 No pre-moderation Reviews publish immediately, with reactive flagging only.
- 04 Active attestation An unchecked box, the last step before submitting, confirming honest personal experience.
- 05 Buildings, not people No reviews of named individuals. Name patterns are auto-flagged.
First prototypes and drawbacks
Claude built a landing page in Figma through the Figma MCP server, then a four-step review flow in HTML/CSS shaped by the five principles. I translated the pages back into Figma so I could make small changes without spending tokens on every tweak.
The translation wasn't clean. Components shifted, icons and the logo went missing, and layers were renamed, all of which needed manual fixes.
Phase 2: A full redesign with a multi-agent workflow
Next I redesigned every screen with DesignPowers, a set of design agent skills that run inside Claude Code. Instead of generating designs in one pass, it runs a structured workflow with me as creative director at each handoff.
The goal was to redesign all screens, add the missing building profile, unify everything in one shared CSS design system, and pass critique, accessibility review, heuristic evaluation and synthetic user testing before calling it done.
- 8screens redesigned
- 1shared CSS file
- 4review agents in parallel
- 4synthetic personas tested
Taste calibration
Calm confidence: Airbnb's warmth with Apple's restraint. Source Serif 4 for display, DM Sans for UI, and gold as the rarest accent.
Strategy
The legal principles became design principles, like “Trust is the product, not the disclaimer” and “Usable under pressure is the baseline”.
Plan and build
Ten tasks with accessibility checks, built into eight screens on one shared token system.
Review agents found several valid issues
Three agents reviewed the build in parallel: a design critic checked it against the brief, a heuristic evaluator ran Nielsen's 10 on the review flow, and an accessibility reviewer checked WCAG 2.1 AA on all eight screens.
| Finding | Severity | Fix |
|---|---|---|
| Step 3 Continue could be bypassed by keyboard | Critical | Replaced the link with a real button, enabled only after 3+ ratings |
| Attestation looked like any other form field | Critical | Green border, shield icon and plain-English copy |
| Legal reminder used red error styling | Critical | Changed to amber. Red is for validation errors only. |
| Move-in dates could be silently skipped | Critical | Required before Continue activates |
AI confabulation and frustrations
Testing the prototype myself in the browser, I found six bugs the agents had missed. The biggest one was the Continue button on step 3 of the review form was permanently disabled. I asked Claude to fix it three times, and three times it said it had. I dug into the code and found the real issue and pointed it out to actually get it fixed.
Synthetic user testing
DesignPowers walked through the key flows as four personas: a mobile-first renter, a low-vision user, a non-native English speaker and a keyboard-only user. It surfaced friction I then fixed: a search bar in the landing page hero, a sticky breadcrumb on mobile, “Tenancy details” renamed to “Rental details”, and simpler attestation wording. A final check added reduced-motion support before handoff.
Reflections
-
AI was great for research
The legal deep dive, understanding publisher liability, etc. were discovered via AI. Research through AI is great when "you don't know what you don't know."
-
Multi-agent review helps, somewhat
Critique, heuristics and synthetic testing caught real issues. But the redesign wasn't dramatically better than Phase 1.
-
Real prototypes pay off
Clickable HTML/CSS screens I could poke at caught bugs that design review missed, and would make an engineering handoff much smoother.
-
It doesn't replace real users
Synthetic personas don't produce true surprises. Nothing replaces watching someone fail at a task on a real device.