How I Built This Research Site with AI Assistance

This site is a static Astro blog hosted on Cloudflare. I built it with heavy AI assistance, and this is an honest account of how that went: not a sales pitch for AI, just what actually happened.

What I Set Out to Build

A low-maintenance, Markdown-first site: write a post as a .md file, push it, and let the host rebuild. No server, no database, no CMS. The constraint was that I wanted to spend my time writing, not maintaining infrastructure.

Where AI Helped

• Scaffolding. The Astro content-collection setup, content.config.ts with a glob loader and a Zod schema, is fiddly to remember. AI produced a correct first version quickly.
• The boring glue. RSS feed, sitemap, security headers, the responsive card grid. All of it is well-trodden ground where a good first draft saves real time.
• Design iteration. Describing a look (“dark, glass, subtle aurora”) and getting CSS to react against was faster than starting from a blank file.

Where It Got in the Way

• Confident but stale advice. Some suggestions used older Astro patterns that no longer match current APIs. The build catches these, but only if you actually run it.
• Plausible-looking gaps. The footer once advertised “cookieless analytics” while no analytics script existed. It looked finished; it was not. AI is good at producing things that look complete.
• Knowing when to stop. Left unchecked, it will happily add features you do not need. The discipline of “minimal version first” has to come from you.

What Actually Shipped

A site I can extend by dropping a Markdown file into one folder. The lesson that stuck: AI is excellent at the first 80% and at anything well-documented, but the last 20%, the judgement about what is correct, honest, and worth keeping, is still the human’s job. Treat its output as a strong draft to verify, never as finished work.