Building this blog in the open

This blog had paperwork before it had pages. A discovery doc, a PRD, four ADRs, a delivery plan. I made the AI write all of it up front, then argued with most of it.

The pipeline behind this page is deliberately boring. MDX files in the repo, compiled at build time, frontmatter checked against a schema. Get the metadata wrong and the build fails:

description: s.string().min(10).max(160),
date: s.isodate(),
draft: s.boolean().default(false),

The process around it is the interesting bit. Agents that carry context between sessions so I don't re-explain the project every morning. Review passes that catch what I miss. And a publish pipeline where the AI edits my words but never writes them. If you're reading a post here, I wrote it.

That's what this journal is for. The decisions, the reversals, the stuff that didn't work. Some of it will be embarrassing. That's sort of the point.