Class: Independent deployment · Confidence: Medium · Demo status: Blog-based pattern
You have more rough prompts than polished outputs. "I should write a runbook for this." "I want an RSS digest of these sources." "Let me sketch a diagram of this system." Each one is a 20-minute task that takes an hour once you open the right tools. You want the path from a sentence in your head to a useful artifact on disk to be one message long.
eBourgess documented their personal Hermes workflow in a first-person blog post. The pattern: rough prompts in, structured artifacts out — runbooks, personal notes, RSS curation, diagrams, Telegram outputs. The agent writes to local files (so everything's grep-able and git-able) and pings Telegram when something's ready.
It's not a product. It's a disciplined setup that turns Hermes into a pipeline from intent to saved artifact, with the author's own conventions encoded in prompts and skills.
- One agent, many output channels. Same Hermes instance, different tools: local file output for notes, Telegram for "done" notifications, research tools for digests, a diagram tool for sketches.
- Prompts that always end in a file. "Write this as
~/notes/2026-04-22-X.md" instead of "write this and show me." Forces the artifact-on-disk habit. - Scheduled curation. Cron jobs pulling from RSS or the web and writing markdown to a notes folder, so the morning brief is ready before it's asked for.
You don't need the author's private config. The ingredients are all first-party:
- Daily Briefing for scheduled research output.
- Telegram gateway for ping-on-done.
- A simple file-output habit in your prompts — always end with "save to
~/notes/<filename>.md". - Optional: a small skill that enforces your filename conventions and folder structure.
- Terminal and file tools (built in)
- Web search / research tools
- Telegram gateway (for delivery pings)
- A notes folder with a convention you'll actually follow
- This is a pattern, not a demo. The value is in the habits — file-first outputs, scheduled curation, one agent in front of many tools — not a specific repo to clone.
- The original blog post is concrete enough to adapt; source is below.