From harvest to publish

This tutorial runs the intelligence desk once, end to end: you register a source you care about, harvest it on a schedule, curate the harvested pile into observer reports, dispatch a writer that turns them into a finished briefing, and finish by taking a newsletter draft through the one gate that matters — your click on Send.

The scenario: you want a standing watch on your field — say, a competitor's blog and a YouTube channel covering your market — that turns into publishable briefings without you reading feeds all day.

Prerequisites

• The install is done and Doctor runs green.
• Your Claude subscription is connected — the curation and writing steps are agent runs (see BYO Claude & keys). The harvesting steps cost nothing: no model call, no API key.
• Optional, for the final step: an email-provider account configured for sending, if you want to walk the newsletter gate rather than stop at the site-bound briefing.
• Budget roughly half an hour of elapsed time — most of it waiting on the first scheduled harvest and the writer run.

Step 1 — Register a source

Open Observer from the sidebar and go to the Sources tab.

1. Choose New source and pick the type — RSS, X / Twitter, YouTube, or Email Newsletter.
2. Paste the URL. You do not need the exact feed address: a channel @handle or a site homepage is enough, and the resolver discovers the real feed.
3. Click Test feed.

You should see: the feed's title, its item count, and a few sample headlines — the probe runs before saving, so you never save a source that does not work. Assign a category, then save. None of the built-in source types needs an API key.

Your install ships with a curated starter registry of AI and tech sources in ready-made batches — this walk assumes you add your own on top, but the starter batches work for every step below too.

Step 2 — Group into a batch and switch on the schedule

Go to the Batches tab.

1. Group your related sources into one batch — the competitor's blog, X handle, and YouTube channel belong together.
2. Give the batch an interval: anywhere from every 15 minutes to every 24 hours, a daily at-a-time run, or a cron expression.
3. Switch the schedule on. Schedules are off on a fresh install — harvesting starts when you decide, not behind your back.

You should see: a warm-up harvest fire about a minute after you enable the schedule. From here the engine runs unattended — it catches up on overdue batches after your machine was asleep and records a run entry even when nothing new was found.

Step 3 — Review the harvest

Go to the Harvest Runs tab and open the newest run.

You should see: what came in, what was rejected as a duplicate, and any per-source errors — with the item bodies readable in place. Every item is clean Markdown on your disk. This entire stage was mechanical: fetching, text extraction, and three-layer de-duplication run with no AI call and no token spend.

Step 4 — Curate the pile into observer reports

Go to the AI Processing tab and dispatch a processing run.

This is the first step that spends: the run reads fresh items and scores, categorizes, summarizes, and curates them into dated observer reports — using your own Claude subscription, and only when there is genuinely new material to process.

You should see: a dated observer report appear in the Reports tab when the run finishes. These reports are the mid-layer the writers build on — and worth reading on their own.

Step 5 — Dispatch a briefing writer

Switch to the Intelligence view and open the Generate tab.

1. Find the Briefing dispatch panel.
2. Optionally steer the run with the free-form instructions field — "focus on enterprise implications", "lean toward critique" — which biases story selection and framing without hand-editing anything.
3. Dispatch.

You should see: a running dot on the tab while the writer works. The writer reads the observer reports, drills into the strongest raw items for specifics, and writes a 200–400 word, impact-ranked briefing with complete frontmatter. Two safety properties hold on every run: a consumption ledger means re-running never re-produces the same story, and a writer aborts rather than pad when the material is too thin for an honest piece.

Step 6 — Decide where site-bound content lands

Open the output directories overlay from the Intelligence view header.

Point the content site root at your website's content folder and every finished piece is picked up by your site's content loader — publishing to your own site is automatic because those are your files, on your disk. If you want to read before anything goes live, point the briefing content type at a private folder instead; pieces then stay on disk, yours only, until you move them.

You should see: the finished briefing as a Markdown file in the folder you chose, carrying an AI-disclosure tag in its frontmatter.

Step 7 — The send gate (the explicit approval step)

Content addressed to other people is a different matter, and this step exists to show you the gate. Dispatch a Newsletter run from the Generate tab.

1. The writer produces a bilingual email digest — the second language written natively, not machine-translated — into your outbox as a draft.
2. Review it. Edit it if you like.
3. Click Send.

Nothing was sent until step 3, and there is no autonomous send mode — the newsletter gate is architectural, not a setting. The send step stamps the file with the send time and the provider's broadcast record, so every issue has an audit trail.

Where everything landed

Make it a standing publication

Everything you just did manually also runs on schedules: the breaking ticker every half hour, briefings each morning, roundups weekly. Enable the corresponding routines and the desk becomes a standing publication — with the same properties you just verified: routines are token-free when idle and only spawn a writer when there is genuinely new material, and the newsletter gate never opens by itself. See Routines & automation for schedules, caps, and retries.

Where to go deeper

• Observer — harvesting — source types, batch mechanics, the newsletter inbox, and video analysis
• Intelligence — publishing — all six content formats, images, and the audit trail
• Routines & automation — the scheduled layer you just previewed manually
• Bring your own proxy — optional, for high-volume YouTube monitoring