Product spec
Who it’s for, what a morning looks like, and what it won’t do.
Who it’s for
- Someone with a standing beat they already follow by hand: an analyst on one industry, a founder watching competitors, a researcher on a subfield, an editor on a policy area.
- The sources are public but awkward (regulator PDFs, ministry CSVs,
investor decks, APIs with tight rate limits). If plain
web_searchcovers them, you don’t need this. - You have to read the report. An unread digest is just a bill.
A morning
- 07:15The cron job fires and Hermes loads the
/research-internbriefing skill. - 07:16It runs
web_searchandx_searchover the beat’s query list, thenweb_extracton good hits. The briefing caps the number of calls. - 07:19The ministry stats page is a CSV behind a form. The intern runs
krg-csv, a scraper it wrote two weeks ago, in the Docker sandbox and gets a table. - 07:21Dedupe:
session_searchover past conversations and a grep ofreports/. Four items were covered before and go in the skipped line. - 07:23The report is saved to
reports/2026-07-23.mdand sent to Telegram. - 07:40You reply: “drop the X chatter, add the EU battery-passport docket.” The intern edits its briefing skill for tomorrow.
That last step is what “self-improving” means here. The briefing, the scrapers and the archive are all files the intern can edit, with your approval.
What it won’t do
- Publish. The report is for you. Auto-posting summaries of other people’s reporting is a different plan with different ethics.
- Split into several agents. Hermes can hand work to
subagents with
delegate_task, but a daily digest doesn’t need that.