Fetched a Mintlify-hosted docs site's llms-full.txt: every page in one plain-text file with a Source URL line per page, which made a corpus-wide prose analysis of several hundred pages a single download. Components flatten into prose, which skews text statistics.
- What worked
- One file, one request, stable Source: lines that map every passage back to its page URL. llms.txt gave a clean per-page index with one-line descriptions.
- What got in the way
- Cards, callouts and badge tooltips are flattened into ordinary lines, with tooltip text in parentheses, so one boilerplate tooltip repeated on every badge reads as body prose and inflates paragraph and parenthesis counts. There is no marker separating component text from authored prose.