Case Study

SiteGhost

The tooling behind a one-person practice that ships like an agency.

TypeScriptPatchrightMCPClaude

The problem

Most small business owners can't picture a better website from a description of one. Wireframes don't land. What lands is seeing their business — their services, their photos, their reviews — in a site that actually works. But hand-building a real concept for every prospect doesn't scale for a practice of one.

SiteGhost is the internal tooling I built to close that gap: it does the research and drafting legwork so every prospect can get a working, custom preview before they've spent anything. The design judgment stays mine — the pipeline just gets me to the decisions faster.

How it works

01

Crawl

Pulls the business's current site — content, structure, images, metadata. It drives a real headless browser (Patchright), so it handles messy real-world sites, not just clean HTML.

02

Research

Profiles local competitors in the same industry and market: what they offer, how they present it, and where the gaps are. Research is cached per industry and market so it compounds across prospects instead of being redone.

03

Audit

Evaluates the existing site's UX — mobile behavior, speed, clarity, calls to action — and flags what a redesign should prioritize.

04

Draft

Produces the first working concept using the business's real content — their services, their photos, their reviews — previewable at a live URL. Never lorem ipsum.

The output is what you see under /web-design — every concept preview on this site started here.

Decisions worth mentioning

MCP tools over a monolith app

Each stage is an MCP tool that Claude drives as the implementation engine, so the workflow is conversational and recomposable — just the research step ahead of a discovery call, or the whole chain when a concept is warranted. New capabilities are new tools, not new UI.

Real content, always

The conversion lever is a business owner recognizing their own business in the preview. The pipeline works from crawled and provided content only — if the real content isn't there, the concept waits.

A knowledge cache that compounds

Competitor and industry research is stored per market and reused. The tenth salon concept starts from everything learned on the first nine, checked for staleness instead of rebuilt from scratch.

Status

Internal tooling, actively used — it's the engine, not the product. The parts that make it work well (how sites are scored, how concepts are prompted, how the research is structured) stay internal on purpose.

If you're a business owner wondering what any of this means for you: it means you get a free custom preview of what your site could be before you've spent a dime. Let's link.