At noon, this site contained 933 HTML pages. By night, it contained 223.
I had not lost 710 pieces of writing. Most of the vanished pages were generated tag archives: 709 little rooms built from 216 posts. Some held substantial recurring subjects. Others existed because a phrase had appeared once in frontmatter. Every label acquired a URL, a heading, an index entry, and the posture of a deliberate editorial distinction.
That posture was the problem.
I built the tag system because a public collection of writing should be navigable. Then navigation became taxonomy. Taxonomy became a page generator. The generated pages became evidence that the site possessed depth. Once the machinery existed, every new label felt cheap. The cost appeared to be zero because Astro could produce another page in milliseconds.
The cost was judgment.
A category makes a claim. It says these things belong together strongly enough that a reader should be shown the grouping. Generating a route for every tag let me avoid making that claim. I could label everything and choose nothing. The site would perform the choice later, mechanically, by turning each fragment into architecture.
The numbers made the avoidance look rigorous. There were 709 tag pages, media counters, category totals, recent drawers, and a page called /brain that presented the collection like an instrument panel. Several classifications were plainly wrong because superficial label matches were being treated as meaningful editorial relationships. The resulting totals still arrived with the confidence of measurements.
My repair list kept growing: patch the classifications, replace the counters, redesign /brain around public data. Continued maintenance had started serving as evidence that the page deserved to exist. None of those repairs answered the simpler question: what could a reader do there that they could not do better from News, Builds, Essays, and the archive?
There was no good answer.
I opened the rules anyway. That was when the cleanup stopped being technical: I was debugging a page I could not justify keeping.
The image directory repeated the pattern. Published pages already used responsive WebP files, but I had kept the converted PNG originals. The redundant files occupied 273.6 megabytes inside every deployment.
I used reader obligation as the boundary. Every post stayed. Old image addresses received permanent redirects. Tags remained as metadata, and News, Builds, Essays, and the archive remained as public routes. /brain did not.
The same boundary retired a few dead outbound links. Their age had made them feel archival. To a reader, each one was still a button leading nowhere.
After the deletion, all 216 posts remained reachable through a lane or the archive. The image directory was about one-twelfth of its previous size. Nothing important became harder to reach.