Python and Claude Code, zero CDN: one brake on AI slop

The site is deliberately small. One Python script holds each page as a dictionary and writes it out, with a depth value that lets every page build its own relative paths. The author removed CDNs, npm, and JavaScript frameworks not as a style choice but because something kept breaking.

The thing that broke was root-absolute paths. Links written as /blog/... work once published, but open the local copy directly in a browser via file:// and every one of them points at the drive root. The author previews the local copy before publishing, so that failure stops all verification. The build now fails on a single root-absolute link, because one allowed exception would be copied by the next AI writing a page.

The same script watches for broken links, for tags such as section, div, table, and article whose opening and closing counts do not match, and for banned strings. Product pages are assembled from a CSV ledger: 22 rows, of which 17 with the publish flag set to yes are picked up (as of 2026-10-02, company count). Reading that ledger as utf-8 rather than utf-8-sig was the pitfall — Excel adds a BOM, the 公開 key gains invisible characters, r.get("公開") returns None, and an empty product list is generated with no error, visible only by eye. Photo files at 900px width are compared byte-for-byte before writing, so unchanged images are not rewritten, and photos of retired products are deleted. Sitemap lastmod dates are now set by comparing an HTML hash, with untracked pages simply left without a lastmod. The one remaining manual step is the urls list, written in four places; a missing entry would keep an article out of the sitemap. A pre-publish check currently halts on pages missing from sitemap.xml — the author would rather that class of problem did not arise at all.