Directing an AI agent through a production rebuild

What this is

I rebuilt back2prod.com (a working music studio's live site) by directing an AI coding agent through 25 days of design, code, and deployment decisions. No engineer. Static site on Netlify, migrated off Squarespace, email continuity preserved, $0/yr hosting.

The site isn't the interesting part. The interesting part is what the process taught me about where AI is leverage and where it isn't.

Three things I learned that I didn't expect

1. The model's quality is bounded by mine.

The agent caught nothing I didn't first catch myself. It duplicated a closing tag and I noticed. It gave wrong-confidence advice about Netlify form detection and I had to redirect. It once over-applied a binaural disclaimer to a track that was natively binaural (a domain-knowledge error I had to correct). None of these would have been caught by a less attentive collaborator. The output ceiling is the input ceiling, and the inputs are driven by human curiosity and judgement.

2. Specification is the actual design work.

"Make it pop" produces garbage. "Bump the home sphere's ambient drift multiplier about 20%, leave other dark spheres as they are" produced exactly what I wanted. The skill that mattered most over 25 days wasn't visual taste or copywriting, it was learning to describe changes with enough precision that the model had only one reasonable interpretation. That's a design skill that wasn't on my résumé before this project and is now the one I'd put first.

3. The honest disclosures are where the model's defaults fail.

The agent's instinct on legally-sensitive copy (AI image provenance, Dolby Atmos vs binaural distinctions, what counts as misleading marketing) consistently softened. It would write "Composed in Dolby Atmos" because it scanned as confident and brand-aligned, missing that it was the third Atmos mention on the page and would read as overclaim. Catching this required holding the whole page in my head and applying a standard the model couldn't infer from local context. The places where the model is most fluent are also the places it's least careful.

What I'd want a hiring team to take from this

The migration is freelance work. The methodology is the portable thing, and the part I'd want to talk about in an interview. Specifically: where in the loop a human's judgment is irreducible, what specification looks like as a deliverable, and how to test for regressions when the agent's small mistakes don't trigger any of the usual signals.

Timeline

25 days end-to-end, with most of the active building concentrated into roughly a week of sessions in late May and another week in early June. I wasn't working on it every day, this fit in around a full time job, life, and some other projects like reset and groovelock. The final two days were deployment, DNS migration, and SSL provisioning.

Outcome

Site live, Squarespace cancelled, email continuity preserved through DNS migration. The full project is at back2prod.com.