Shanelle Llamido — Product, AI Website Builder
Work sample — Product Manager, AI Department

What if publishing a website took one sentence?

This page is the work sample. The workbench below is a working, simulated version of the product's core loop — describe a business, watch it become a live page. Everything after it is the product case study behind that loop: who it's for, how it should be built, what to measure, and where it can go wrong.

A product case study by Shanelle Llamido, prepared for the AI Website Builder role.


The workbench — live simulated loop

Describe a business. Watch it get published.

A rule-based generator maps your description to an industry template, drafts copy, applies a theme, and "publishes" a page — while a product telemetry panel logs every step, the way a real build would need to be instrumented.

Tone
artisan-coffee-roastery.site
Why this, why now

The gap was never power. It was the first ten minutes.

Website builders never had a power problem. Wix, Squarespace, and WordPress can already build almost anything a small business needs. What they never solved is the first ten minutes: the blank canvas, the seventy-template carousel, the slow realization that a decent site is a lost weekend. Most people close the tab. The ones who push through tend to produce something that looks like every other site built from the same starting point.

AI collapses those ten minutes into one sentence and a few seconds. But the hard part isn't the model — it's a product decision. Ship something fast and generic enough and every AI-built site converges on the same three layouts, the same voice, the same stock structure. Speed without a point of view is just a faster template. The job is to make the generated draft a starting point people actually want to edit, not a ceiling they have to escape.


Who's on the other end of the prompt

Three people, three definitions of "done."

Archetype 01

The Side-Hustler

Launching a thing this week around a full-time job. Has a name, a logo maybe, and forty spare minutes tonight. Judges the product on whether there's a real URL by bedtime.

"Good enough, published" beats "perfect, still in draft."

Archetype 02

The Local Business Owner

Runs a salon, a clinic, a bistro. Low tech confidence, high stakes — this site represents the business. Won't open dev tools, won't debug a broken layout, will just leave if it feels fragile.

Needs to trust it before they'll ship it.

Archetype 03

The Independent Creator

A designer, writer, or consultant who will keep tweaking for weeks. Cares about every margin. The real stress test of whether editing an AI draft feels good or feels like a fight.

The true test of the editing experience.


Product principles

Four rules I'd hold the roadmap to.

01

Speed earns the first look. Quality earns the domain purchase.

A fast generation buys thirty seconds of attention. The domain purchase depends on whether the draft survives a careful read — so optimize the whole funnel, not the first screen.

02

Regenerating a section should never feel like starting over.

Section-level control, not full-layout reshuffles. If editing the hero also shifts the pricing table, people stop trusting the regenerate button — and that button is the product.

03

AI confidence should never outrun AI accuracy.

Factual fields — hours, pricing, credentials — need a visible review-before-publish state. Confident, wrong, and public is the failure that costs a real business a real customer.

The workbench above demonstrates this: every draft carries a review-before-publishing badge.
04

Great defaults, real editing underneath.

The generated page should be good enough to ship untouched — and every part of it editable when someone wants to. Defaults are a starting point, never a cage.


Roadmap

Prove the loop, then earn the right to expand it.

Sequenced so that trust and control land before growth — a bigger funnel into a product people don't trust just scales the churn.

Phase 0

Prove the loop

One template family, one path: prompt in, one live page out. Tone quality checked by hand on every generation while we learn what "good" sounds like.

1 template family prompt → 1 live page publish rate = north star
Phase 1

Expand coverage

Enough industry range that most visitors see something that fits, plus the first real editing affordance: regenerate one section at a time.

8–10 industry templates brand kit from uploaded logo section-level regeneration
Phase 2

Earn trust & control

Make editing an AI draft feel safe. Nothing gets silently overwritten; nothing ships that fails a basic quality bar.

inline editing undo history pre-publish a11y / SEO checks
Phase 3

Growth loop

Turn published sites into the acquisition channel — real domains, real analytics, and a gallery that shows the range of what people actually built.

custom domains owner-facing analytics remix gallery of real sites

How I'd run it

The decisions I see coming — and how I'd approach them.

Four forks the PM of this team hits in the first two quarters. For each one: the options on the table, where I'd land and why, how we'd know it's working, and what would make me reverse it. Click any option to compare.


Target metrics I'd watch from day one

Five numbers that tell you if the loop is working.

Illustrative targets to instrument and steer toward — not live data from a running product.

Time to first publish <5 min Prompt to a live URL, first session.
Activation rate ≥55% Signups who publish at least one page.
Edit-to-publish ratio <3 Edits needed before someone ships. Lower = better defaults.
Regeneration rate <20% Sections regenerated. A quality-miss signal, not engagement.
Day-7 return rate ≥30% Owners back within a week to edit or extend.

Competitive landscape

Everyone can generate. Few make editing the draft feel safe.

ApproachStrengthWhere the gap is
Wix ADI / Squarespace Blueprint Mature, deep editor once you're in; huge component libraries. Still template-bound — the AI picks a template, it doesn't change what a site can be.
Framer AI / Durable Fast, AI-first generation; a live site in under a minute. Editing after the fact is thin — you regenerate or you fight the canvas.
10Web / Dorik AI Real power underneath via WordPress and full builders. That power comes back as complexity — the first-ten-minutes problem returns.
The opportunity Not a smarter generator — a safer one. Make regeneration feel safe, so editing an AI draft feels like refining, not gambling.

Risks I'd de-risk early

The failure modes that don't announce themselves.

Hallucinated business claims — invented hours, prices, awards, or credentials published as fact.
Flag every factual field for explicit confirmation before publish; never let a generated claim ship unreviewed.
Sameness at scale — thousands of AI sites converging on three layouts and one voice.
Track template and layout diversity per category as a product health metric, with alerts when it collapses.
Accessibility and SEO regressions — generated markup that looks fine and scores badly.
Automated pre-publish audit (contrast, headings, alt text, meta, structured data) that blocks or warns before going live.
Trust erosion from silent overwrites — a regenerate wipes an edit someone spent time on.
Undo history from day one, not a v3 feature. Every regenerate is reversible and visibly so.

Closing note

Why a demo instead of a follow-up email.

A status-check email tells you I'm still interested. A working version of the core loop — instrumented, with the trust cues built in — tells you how I'd actually approach the role: start from the user's first ten minutes, make the risky part safe, and measure whether it's working. That felt like the more honest thing to send.

Shanelle Llamido
Re: Following up on the Product Manager, AI Department application
Email shanelle.llamido@gmail.com