Using data to build with AI

Building it end to end

One person and one AI model built all of it: design, code, email, delivery, the search pages, the tracking. Next.js and TypeScript on Vercel. gpt-image-2 draws the card. GPT-4o writes the prompt. Resend sends the email, Vercel Blob holds the share and gift links, Airtable takes the leads, Upstash runs the rate limits, and pdf-lib makes the printable.

Claude Code runs the codebase under a doc system: a briefing, a state file rewritten every session, a session log, a learnings file, and an archive that fills before anything is deleted. More than two hundred and fifty sessions so far. Every change passes the same gates before I see it: types, lint, the test suite, the build.

• Next.js
• TypeScript
• Vercel
• OpenAI
• Gemini
• Resend
• Airtable
• Upstash
• Claude Code

The loop

The app sends an event for every step a person takes. Vercel keeps those events. Google Search Console and Bing Webmaster hold the search side. Airtable holds the leads.

Once a week Claude Code pulls all of it into a dated snapshot, diffs it against the week before, and writes a short memo. I read the memo and decide what changes. Claude builds the change, verifies it, and hands me a test list for my phone. This site reads the same pipe when it builds. Nothing above was typed in.

Instrument before you grow

The tracking went in before the first search visitor arrived. About sixty custom events, written so a finished card is counted apart from a click and a delivered email apart from a sent one. Test surfaces carry a marker and every query filters them out.

Definitions came before rates. On a small site the owner is the traffic, so every read asks the same question first: was any of this me? The homepage conversion rate once counted people who had walked in from another page. The fix was a range with the pass-through subtracted, not a bigger number.

Connected sources

Vercel Web Analytics is read through its MCP server. Search Console through a service account. Bing through its API. Airtable holds the leads. A skill runs the weekly ritual in one pass: pull, snapshot, diff, memo. A rate built on fewer than twenty visitors is called unreadable rather than quoted.

In the last 30 days 93 of 449 visitors came from a search engine.

What the numbers changed

Three reads from the memos. Each one is a number, a call, and what shipped.

Bing ranks the plain exact-match page at position 4 for its head term. Google puts that same page at 10 and ranks the themed pages instead. The call: never clone the head term. The milestone-age page shipped the next night.

In the 30 days to September 11, search visitors who landed on a themed page made a card 43% of the time. On the homepage it was 19%. The call: the homepage was never the problem. Six pages had no search visitors at all, and that is where the work went.

17 of 330 renders failed in a 31 day window, and one morning the model refused my own photo four times in a row. The call: a render never fails closed. If a guard trips, a fallback concept carrying the typed details runs instead. And no prompt on a photo path ships until eight or more real photos pass at zero refusals.

Working together from the command line

Everything on this page came out of the same terminal: the analytics reads, the copy, the prompts, the code, the docs, and this page itself. I open a session, say what I want in plain words, and Claude Code reads the project's own files before it touches anything.

I ask the question and read the memo on my phone. Every trade-off gets framed by Claude and called by me. The push comes from GitHub Desktop, by hand, and Claude never commits. The AI does as much as it can. The direction and the thinking behind it are mine.

The same model on a business

A business already has the sources: orders, web, ads, support, finance. Connect them into one warehouse and fix the definitions before the first dashboard. Then let AI read the pipe, draft the read, and build what gets picked, with a person deciding and owning every change.

Questions