A scrape, a prompt and a digest

The team watches about forty pages: regulators, two competitors' changelogs, a handful of forums, a few people's feeds. Someone used to open them all on a Monday. By Thursday nobody had.

The scrape is an Apify actor with the forty URLs, scheduled for six. It lands in Emidox as a dataset that grows by one entry a day. The agent's prompt asks for one thing: compare today's copy of each page with yesterday's, and list only what changed, grouped by why it matters to us — and it is given the "why it matters" as a paragraph about what the team does. The model is Gemini Flash, because reading forty pages is a volume job and a cheap model does it well.

The output is a Telegram channel the team already had. At 6:30 there is a message with the six or so changes that matter, each with the page it came from, and a line at the bottom: "34 pages unchanged". Nobody opens forty tabs. When something is wrong — it once flagged a date format change as a policy change — the fix is a sentence in the prompt.

The part worth copying is not the scrape or the model. It is the phrase "only what changed". Most digests fail because they summarise everything, and a summary of everything is just a shorter version of the reading you were trying to avoid.