Builder Series 01: How a Solo Founder Ships Her Entire Product and Catches AI Slop

Why Catherine Built Annsa

Cat spent twenty years in strategy at WPP, commercial growth at TradeGecko, and product at Intuit across 220-plus markets. No matter the role, the first move was always the same: go find customer feedback. “It didn’t matter what tools I used — enterprise-grade or hacky sort of stuff — I’ve just never really been happy with it.” She started her career as a researcher, and that’s the thread running through everything since: you always do things to solve customer problems.

That’s what became Annsa — her product now, after a few names and pivots along the way (it started as “In Bloom,” then “Circuit,” before landing on Annsa). It pulls together unstructured customer feedback, sorts and ranks it, and turns it into a roadmap grounded in your actual codebase. The job has stayed the same the whole way through: work out what your customers want, and know what to build next.

Designing Without a Designer

Cat isn’t a designer. She’s led marketing design teams before, but calls it “certainly not my strength.” She started in Figma Make — describe the product, let it build. About a thousand chat iterations later, she had a working prototype, and showed her husband, thrilled: “Oh my God, it’s amazing, look what I’ve built.” Then the realization set in: it was AI slop. “Everybody’s product looks like this.”

So she went back through everything, component by component — cards, accordions, how integrations work, how feedback data gets added, how it connects to GitHub — every single piece of the experience, rebuilt one at a time, chatting back and forth with Figma Make on how to build and design each one. Every component got eight to ten design options on the table before she picked one. Priority display alone had four: a double-height row, a more mobile-responsive layout, a card grid, and a Kanban-style board grouped by status.

Around the same time — before Claude Code had really taken off — she was running this in Cursor, and Cursor generated 72 versions of a single button. “I had to go and say no, this is the component. This is how we’ll do a checkbox. This is how we’ll do a text area.” Not her strength, by her own account, but doing it by hand gave her real appreciation for everyone who’s done this work before.

Once the component set was locked — a fixed number of buttons, defined sizing worked out — she moved off Figma entirely for the day-to-day. Still loves it, but she designs in Paper now, building from the base component set she’d already fought for.

Four Models, One System

With the foundation in place, Cat split real work across four AI models by role. On the marketing site: Codex drafts the first iteration, Claude critiques it — “sometimes they fight, but they critique each other” — and Grok comes in with another viewpoint. On the product itself, the order flips: Claude first, then feedback from Codex and Grok. All three are hooked into Paper by MCP, alongside Cursor, with the codebase on GitHub so every agent can check its work locally before it comes back to critique.

One night she gave the agents a brief before going out for dinner with old friends, and told them to keep debating until they’d closed off each other’s open questions. She came home to 3,000 lines of agents arguing with each other. Some of it she overrides on feel alone — working on shadowing and accessibility, the agents landed on something she knew was technically correct but just didn’t like: “I don’t like the feel of it, so we’re not gonna do it that way.” Her read on unsupervised AI coding tools: “I would have 2,000 buttons, 2,000 features. Everything would be really busy and overwhelming — it wouldn’t be the right amount for an experience.”

On taste itself, she’s candid: “Some people are born with it. I’m still learning how to get that taste — it’s not a natural thing for me.” The most practical habit she’s found is asking Claude for optical balance — catching layouts that are technically symmetrical or perfectly aligned but don’t feel right, then adjusting until something reads warmer or friendlier. It’s a genuinely useful reframe: using an LLM to catch things that are mathematically correct but visually off, which is exactly the gap most non-designers don’t know they have.

That coordination scales up fast. On a typical morning, Cat counts up to nine terminals running at once across Claude, Codex, Grok, and Cursor — nine different things being worked on in parallel. To keep them from colliding, she’s built logs and ledgers tracking what every agent and sub-agent is doing, plus shared context on the roadmap, what customers are asking for, and the jobs-to-be-done behind each request — so the terminals keep checking their own work back against the brief instead of just each other.

Recap, in her own words: Claude for context, Codex for visuals, Grok to critique, Cursor for end-to-end tests.

She’s also built a habit around failure. After Anthropic’s Boris posted that you should ask Claude for lessons, she started a lessons.md. Then came a second observation — that agents get discouraged if they think they’re never winning — so she added a wins.md alongside it. For two months, every ship recorded both: the lesson if something went wrong, the win if it went right. Too many lessons and not enough wins, and the agents visibly got worse.

Fresh Eyes: The Audit She Wrote Herself

Cat calls her audit skill Fresh Eyes — built because she wanted a genuinely outside perspective on her own screens. It runs across 100-plus screenshots spanning desktop, mobile, and tablet, and scores every surface against a set of principles on a scale from rough to competent to professional. An LLM then reads the raw results and writes them up as a report.

A run from the morning of the interview: Priorities scored competent. The product reader — where she gets commercial guidance on how the business is doing — scored professional. The roadmap Kanban board was still rough. A brand-new “Wild” feature (in-the-wild feedback) scored rough simply because it wasn’t live yet. The first time she ran Fresh Eyes on it, the output was overwhelming — “so many things, you’re like, what do I do” — so she narrowed the ask to just the top three systemic problems. That pass surfaced font-weight inconsistency, dark mode needing weight compensation, and a couple of radius-and-nesting issues, alongside a list of what was already working: fewer transitions for a calmer feel, shadow treatment, timestamp formatting.

The full report goes page by page — IA observations, what’s working, and a verdict framed as “if the experts redrew this, what would they recommend” — down to individual settings sub-pages, plus a final summary across the core surfaces. From there, she sets Claude loose to fix what needs fixing. Not everything gets changed: sometimes a second pass with an LLM judge confirms a flagged issue is actually a deliberate design choice worth keeping as-is; other times, it’s a clear “go fix that.”

Cat’s blunt about where this started: “When I started, everything was rough.” Getting from a 0% baseline to 50–70% happens fast. Every point past that is “experiment after experiment after experiment” — the same grinding iteration she used to build the frontend, the backend, and the end-to-end test suite in the first place.

The Catch

One story sums up why a human still has to be in the loop. That same morning, the team was working on a row hover state — making it more visible, checking it held up for accessibility. Somewhere in that pass, an agent added a colored stripe running down the left edge of the hover. “AI slop central,” Cat called it. “I was like, no — how did that happen?” She caught it, they turned it off, and everything else from that batch shipped as planned.

Nine things shipped correctly. One thing read as AI-generated. She was the only person positioned to notice.

Her take on why that catch has to be human: as a solo founder, “you’ve got a team of 100 people with these agents,” and there’s no version of that where one person has the brain space to be across everything. The only real lever is making very good decisions and building very good systems — not trying to personally review every output. That’s exactly why she’s building a product intelligence system in the first place: if customer feedback can reach the same scale and reliability that her coding agents already give her, “it makes everyone really quite powerful.”

Where Designers Fit Now

What stands out across this whole conversation is that Catherine never skipped design — she took it seriously. She built a component library before she built the product. She tested early, went back, and refined. She’s still working on taste, and she said so herself, out loud. That’s a designer’s process, running end to end.

These tools didn’t make design optional. They made it unavoidable. Every one of those calls — which of eight options to ship, whether a technically-correct layout feels right, whether a hover state reads as slop — still has to be made by a human with judgment. Catherine is making all of those calls by herself, and she’d be the first to tell you it’s an enormous job.

If you’re a designer watching this, the takeaway isn’t that founders don’t need you. It’s that a lot more people are now doing design work, and they’re hitting the exact same walls designers have spent years learning to navigate. That’s a bigger audience for what you know — not a smaller one.

Tools & Resources

Everything mentioned in this one, free to use:

• Annsa — Cat’s product. Turns unstructured customer feedback into a ranked roadmap, grounded in your codebase.
• The Fresh Series — Catherine’s three outside-in audit skills (design, jobs, growth), judged against named masters rather than the product’s own rulebook. MIT-licensed.
• DESIGN.md — a format spec from Google Labs for describing your visual identity to coding agents: tokens, plus the reasoning behind them.
• Impeccable by Paul Bakaus — 23 commands and 59 detector rules for AI-generated frontend design. Its detector catches the exact side-tab-border slop Cat spots in this episode.
• Skills for Designers and Engineers by Emil Kowalski (ex-Vercel, ex-Linear) — animation and design skills built from real domain expertise.
• Refero — a design reference library.
• Paper — where Cat designs now.
• Prettier — opinionated code formatter.

If you’re a founder or small team trying to set up a design system your AI agents can actually work inside — components, rules, and a way to check the output — feel free to book a free consultation. No pressure either way, happy to have a chat and get to know your team.