The Bleeding Edge

// Building with AI · Live

Research Content Pipeline

The system that writes this site. Every Friday it reads 18 AI newsletters and turns them into a sourced briefing, articles and three newsletters. Every morning it writes the daily pulse.

View KPIs →

We read about 18 AI newsletters a week so our listeners don't have to. This pipeline does the reading first.

Every Friday at 04:00 it collects the week's newsletters, merges duplicate stories, and has Claude write the episode briefing. Every story gets its sources and a trust label. From the same material it writes deep-dive articles, three newsletters (LLM Weekly, Executive Roundup, and Devices & Robotics) and the AI Timeline update, then publishes everything to this site. Separate daily runs write the morning pulse and the industry desk.

Everything on this site under Episodes, Articles, Industry and the homepage pulse came out of it. That's the proof, and you can read all of it.

// Who it's for

  • ▸ Newsletter writers, podcasters and analysts who read more than they can keep up with
  • ▸ Teams that brief leadership on a fast-moving field every week
  • ▸ Anyone who wants AI-written research with sources and trust labels on every story

// Who it isn't for

  • ▸ Anyone after a hands-off content farm. Every story carries its sources so a person can check it.
  • ▸ Anyone expecting a hosted app today. There's no SaaS; the done-for-you option is a waitlist.

// Proof it works

// Build it yourself, build it with us, or have it built

Build it yourself

Starter prompts

Free

You can already build. Copy these and go. No email needed.

Newsletters → sourced story list

Paste one or more newsletters. You get a de-duplicated story list with a trust label and the original links on every story, ready to write a briefing from. Replace the {{PLACEHOLDERS}} first.

You are the research producer for a weekly news briefing. I will paste one or more newsletters below. Turn them into a clean, de-duplicated story list I can build a briefing from.

For each distinct story:
1. **Headline**: one line, specific (name the company, product, amount or date).
2. **What happened**: 1-2 sentences, facts only.
3. **Why it matters**: 1 sentence for {{AUDIENCE}}.
4. **Label**: exactly one of
   - [Corroborated]: two or more independent sources in the text, or a primary source (the company's own announcement, a filing, a paper)
   - [Unverified]: a single secondary source, a leak, a social post, or unnamed sources
   - [Inference]: your own reading, not a reported fact
5. **Source**: the URL(s) exactly as they appear in the newsletter. Never invent or "fix" a URL. If the newsletter gives no link, write "no link in source".

Rules:
- Merge duplicates: if two newsletters cover the same story, list it once and keep both links.
- Keep opinions out of the fact lines. If you have a view, put it in a separate line labelled [Inference].
- Drop sponsored items and ads. List them at the end under "Skipped (sponsored)", headline only.
- Order the stories by how much they matter to {{AUDIENCE}}, most important first.
- End with "Top 5 this week": the five headlines you would lead with, one line each.

Newsletters:
{{PASTE NEWSLETTERS HERE}}

Tested on Claude Opus 5.5 on 28 September 2026, with our own LLM Weekly W39 newsletter as the input. It returned a labelled story list; all 9 links it cited were in the input, and it flagged a date conflict in one story instead of smoothing it over.

Build it yourself

Pro pack

$29

Builders who want a production-quality build and a deploy runbook.

  • Every prompt the pipeline runs: weekly briefing, deep-dive article, three newsletter formats, daily pulse, industry desk, timeline extractor
  • A starter repository with the config, folder layout and the de-duplication rules
  • Deploy runbook: scheduled runs on GitHub Actions, a static site on Vercel

Planned contents. Details may change before release.

Coming soon

This also subscribes you to our free weekly newsletter. Unsubscribe any time.

Build it with us

Guided build

$49

You can't build yet. Every step, every account, every command. Pro pack included.

  • Everything in the pro pack
  • Step by step from zero: Claude Code, GitHub, Vercel and a dedicated mailbox
  • Every command, with what it should print
  • What it costs to run, measured on our own build

Planned contents. Details may change before release.

Coming soon

This also subscribes you to our free weekly newsletter. Unsubscribe any time.

Have it built

Have it built

You want the outcome, not the build.

  • We set it up on your sources, your categories and your site
  • Your voice and your schedule, with sources and trust labels on every story

Want us to build it? Join the waitlist

This also subscribes you to our free weekly newsletter. Unsubscribe any time.

Paid subscribers of The Bleeding Edge will get 25% off packs and guides once paid subscriptions open. Subscribe free on Substack →

// Rules and limits

  • Newsletters are other people's work. The pipeline summarises and links; it doesn't republish anyone's text.
  • AI output can be wrong. The trust labels tell readers how far to rely on each story. Our site publishes automatically, and we've had to correct mistakes after they went live. If your name is on it, read it before it goes out.
  • It reads an inbox. Use a dedicated mailbox with read-only access, not your personal email.

// Build log

  1. 2026-09-28KPIs page shows real figures only; anything not yet measured says soe9fc21a
  2. 2026-09-28Explainer cards live, drafted from the site's own sourcesd9e93b3
  3. 2026-09-27Site moves to bleeding-edge-ai.com25709cd
  4. 2026-08-12Six-sector industry deskc0d84cc
  5. 2026-07-11Fix: briefings had been coming back as chat summaries; tools are now off for every pipeline callb4f9263
  6. 2026-07-05Daily pulse and the living AI Timeline8e8f571
  7. 2026-07-02Newsletter ingestion moves to IMAP, ending weekly login failures81c2afe
  8. 2026-05-09The Friday pipeline goes live: first automated briefing, article and three newsletters (W19)cffdbf9
  9. 2026-03-23First commit

// Case studies