Four apps, eight weeks, one designer

Web3Connect is a vertical reviews-and-recommendations marketplace for the blockchain industry — conceptually G2 or Clutch, but specific to one sector. Service seekers find and compare industry partners; partners claim and upgrade their listings; reviewers leave structured, verified feedback.

It was also a deliberate, time-boxed experiment: could one designer with AI as the engineering team ship a real, quality product? The bar I set was that no flow could be an easy-path default. Every key surface had to be thought through on its own terms: product, engineering, UX flow, customer lifecycle stage.

Five production surfaces, all responsive web, all mine end to end:

I inverted the usual order. The design system was written in code first — shadcn/ui and Tailwind v4 as primitives, in a single shared component package consumed by all four apps — before I drew a single screen.

The reason is specific to working with an AI engineer. The agent reads the system as it builds. If the source of truth is a Figma library, the agent is working from my description of the system; if the source of truth is the component package, the agent is working from the system itself. A change to a token or a component propagates globally instead of requiring me to notice a drift and correct it.

A two-sided marketplace has a hard onboarding problem: the same URL has to serve people who want opposite things. I built the dashboard around three archetypes — service seeker, offering reviewer, and partner (free versus paid tier) — and made the interface adapt rather than ask.

• Banners, menu options and primary calls to action all shift with archetype and tier.
• Users are tagged from behaviour during registration paths and key button clicks, and that tagging shifts downstream messaging — including email.
• The surface always shows a next-best action: leave a review, claim a listing, upgrade tier, buy a visibility boost.

The full write-up will cover the state matrix, what happens when someone is legitimately two archetypes at once, and where the adaptation was too clever and had to be dialled back.

A growing share of traffic to a discovery-and-comparison site arrives as an LLM acting for a user. For that visitor, information structure and density are the visual hierarchy — there is no visual layer.

So I treated the agent as a user archetype with its own interface:

• A /llms.txt index, so an agent can find the shape of the site without crawling it
• Markdown responses instead of HTML where the consumer is an agent
• X-Markdown-Tokens headers so an agent can budget its context before fetching

This is the part of the project I find most interesting, and it’s an active research interest rather than a solved problem — see what I’m researching.

The failure mode of AI-assisted building is drift: the output is plausible, ships, and is quietly worse than what you’d have accepted from a person. My answer was to encode the standard as rails rather than review for it after the fact — a modular instruction system of 15+ instruction files and slash commands covering design conventions, accessibility rules and code-review standards.

Method was a markdown PRD, a polymorphic schema, taxonomy and roadmap, worked through sprint by sprint. Research came from PostHog session replay and custom event capture, which is how I found the confusion points that turned into more completed flows and less downstream admin.

Post-launch I read the macro signals against further investment: AI agent chat displacing traditional search as the discovery surface for service seekers, declining attention on the crypto industry generally, and a Clutch-style top-visibility business model facing real competitive headwinds.

So I stopped investing in it. The platform runs on auto-pilot, and I’m redirecting my full-time effort into a team that can use the experience. I’d rather include this than leave it out — calling time on your own product is a design and commercial judgement too, and I think it’s the more useful thing to know about me.