Intelligence — publishing

The Intelligence view is the production half of the intelligence desk. Where Observer harvests and de-duplicates, this view turns the clean pile into finished, on-brand, source-grounded content — and publishes it to a content folder you choose, typically your own site.

The whole desk in one line: sources → items (de-duplicated) → observer reports → writers → your site and your outbox → your approval for anything that reaches another person.

Three tabs — Overview, Generate, Images — plus an output directories button in the header that configures where each content type is published.

The six content formats

Each writer is a fresh, disposable agent spawned for one run; the schedule, de-duplication, file paths, and send step are all mechanical.

Briefings carry an impact tier — notable · significant · major · critical — computed from how many independent sources reported the same story. The story three sources covered outranks the single-source one, mechanically.

Dispatch a writer

1. Open Generate. Each format has its own dispatch panel, and a running dot on the tab shows any writer currently in flight.
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 editing anything by hand.
3. Dispatch. The writer reads the observer reports (and drills into the strongest raw items for specifics), writes a finished piece with complete frontmatter, and exits.

Two properties keep repeated runs safe. A consumption ledger means re-running a writer never re-produces the same story — material already turned into a briefing is skipped, and a force-reprocess run explicitly asks for a new angle or aborts rather than duplicate. And writers are grounded: when the source material is too thin to support an honest piece, the writer aborts instead of padding.

Most formats also exist as scheduled routines, so the desk can run as a standing publication — the breaking ticker every half hour, briefings each morning, the roundup weekly. Schedules, caps, and source filters are all editable; see Routines & automation. Writer runs use your own Claude subscription, and the always-on routines are token-free when idle — they only spawn a writer when there is genuinely new material.

Publish to your own site

The header's output-directories overlay holds the one setting that decides where content goes: the content site root, plus per-content-type overrides. Point it at your website's content folder and every finished piece is picked up by your site's content loader; point a type somewhere private and it stays on disk, yours only. Newsletters never publish to the site — they are email-only by design.

Voice and language come from your identity files: your brand voice, byline, audiences, and publishing locales are configuration every writer reads. Fill them in once and the whole desk produces in your voice, in your languages — the bilingual newsletter's second language is written natively, not machine-translated.

Images

The Images tab manages the illustration layer: hero images for briefings, reports, and roundups, generated from prompts composed out of the full article text. The image service de-duplicates against images already on disk, patches the image into the article automatically, keeps a cost ledger, and enforces a daily cap. Image generation uses an image-model API key you bring — it is optional, and everything else on the desk works without it.

The human gate and the audit trail

Publishing to your own site is automatic — those are your files, on your disk. Anything addressed to other people is not:

• Newsletters are always drafts. The writer produces them into your outbox; nothing is sent until you review and click Send. There is no autonomous send mode.
• Every send is recorded. The send step stamps the file with the send time and the provider's broadcast record — a per-issue audit trail. Sending uses your own email-provider account, configured once.
• Every published piece carries an AI disclosure. Writers set a disclosure tag in the article frontmatter, and a site that renders it can display a locale-aware disclosure automatically — the posture the EU AI Act expects, enforced per article rather than promised in prose.

The Reports viewer

The report archive lives in the Observer view's Reports tab: the dated observer reports the processing layer produces, plus the deep video-analysis reports. This is the mid-layer worth reading directly — writers build on it, but each report is a useful intelligence artifact on its own, and the viewer lets you browse the archive by date without leaving the app.

What this desk deliberately does not do

• It does not invent facts — thin material aborts a run rather than padding it.
• It does not send anything to other people without your click.
• It does not machine-translate — translations are written in-agent for tone, which trades raw speed for quality.

Where to go next

• Observer — harvesting — the intake half of the desk
• Routines & automation — schedules, retries, and where you stay in control
• The operator model — the approval gates the whole platform enforces