SDSignal Desk
Human gate on every publish

One brief. Three formats. Nothing ships unread.

Signal Desk researches your beats, picks one angle worth publishing, and produces a LinkedIn post, a carousel and a newsletter issue that all say the same thing. Then it stops and waits for you.

11
machine gates
3
revision rounds, then it parks
0
posts published without you
2
measurement windows
LinkedIn post
Hook engineered to land inside the 210-character truncation, capped at 3,000, with the evidence attached.
Carousel
A real PDF document post, rendered from structured slide JSON rather than generated as pictures of text.
Newsletter
Sent by Resend, plus a clean payload for the LinkedIn editor, which has no API.

It is a loop, not a pipeline

A pipeline produces the same quality on day 300 as on day one. The difference here is the last arrow: what performed changes what gets picked next.

  1. 01

    Harvest

    Firecrawl sweeps your beats and dedupes against 180 days of history.

  2. 02

    Curate

    A cheap model ranks relevance, novelty, authority and recency.

  3. 03

    Plan

    One strong call picks one angle and writes one brief, every proof point bound to a source.

  4. 04

    Draft

    Post, carousel and newsletter written in parallel from that single brief.

  5. 05

    Render

    Slide JSON to React template to a measured, overflow-checked PDF.

  6. 06

    Verify

    Eleven machine gates, then a critic that never sees the writer's reasoning.

  7. 07

    Review

    You approve, edit or reject with a reason. Nothing ships without you.

  8. 08

    Observe

    T+24h and T+72h metrics feed back into what gets picked next.

What it will not do

  • ·Publish anything you have not read and approved.
  • ·Pretend LinkedIn has a newsletter API. It does not, so you paste that one.
  • ·Generate slides as pictures of text, which cannot be edited or kept on brand.
  • ·Keep spending after the per-run budget is reached, even to finish a run.