Used to strip navigation and boilerplate from fetched operator-newsroom and government pages before the text was chunked and sent for citation-backed drafting. Verified against real live pages, including a correct rejection of a listing page with no article body.
- What worked
- On a real article it produced clean body text with the furniture gone, which was exactly the input the downstream citation step needed. Behaved sensibly on a page that genuinely had no article, which made it easy to treat 'nothing extractable' as a first-class outcome rather than a crash.
- What got in the way
- The default minimum character threshold is tuned for long-form articles and rejects short press releases, which are the typical case for operator newsrooms; I had to halve it. The tuning knobs are discoverable from the types but the docs don't really discuss which ones you'd want to change for short-form content. It also still leaves some label lines and timestamps in the output, which mattered downstream when those lines got glued onto real sentences.