Six weeks ago I had an idea for an AI tools review site.
I have no coding background. I work full time at an airline.
Today talentedatai.com is live with 21 articles, a working
build pipeline, responsive images, a newsletter, and a
Product Hunt launch. I haven't written a single line of code.
Here's exactly how I built it:
The stack:
How the content pipeline works:
I give Cowork a prompt describing what to write. It researches
the topic, writes a 2,500-word article with front matter, FAQ
schema, internal links, and SEO metadata. It saves the file to
a pending folder. I review it. I approve it through a dashboard
that moves it to published. Cloudflare builds and deploys
automatically on every push.
I haven't touched HTML, CSS, or JavaScript directly. Every fix,
feature, and improvement has been done through conversation.
What surprised me:
What's working so far:
What I'm figuring out:
The experiment I'm most interested in:
Can a site built and operated almost entirely by AI, with no
personal brand behind it, grow organic traffic and eventually
generate meaningful revenue?
I'm treating this as a genuine experiment, not just a side
project. Documenting everything.
Happy to answer questions about the technical setup, the
Cowork workflow, or what I'd do differently.
Site: talentedatai.com
Product Hunt: https://www.producthunt.com/products/talentedatai
The workflow you're describing works until Claude resets context between sessions. Suddenly it forgets the decisions it made last week and starts suggesting patterns that contradict what's already in the codebase.
The fix that stabilized ours: a CLAUDEmd file at the repo root. It's a plain text document that describes the stack, the patterns we use, the patterns we've explicitly ruled out, and why. Every new Claude session starts by reading it.
Without it, Claude is a brilliant intern with amnesia. With it, it's more like a collaborator who read the project brief.
For a non-coder building something real, this changes the risk profile significantly — you stop being surprised by outputs that contradict each other.
How long did it take you to get stable output without it?