In February the Iran story was breaking, and I was too busy building other products to keep up with it. So I asked Claude for a quick rundown each day. It’s very good at that. I pasted the briefings into a few nomad groups I was in, and people loved them. I kept thinking: I’m literally just asking Claude, and anyone could do this.
The trouble is that a prompt starts from scratch every day. It doesn’t remember what it told you yesterday, and nothing checks what it says. The rest of this piece is how Lowdown deals with that.
I keep the prompts, thresholds, source lists and word lists private.
How a briefing is made
Each briefing passes through 21 departments, grouped into eight stages that always run in this order:
- Research checks first-party sources before news outlets: government sites, press releases, changes to web pages and data feeds. It files each finding separately.
- The editorial conference writes a brief: the lead, the angle, the counter-angle and what is left out.
- Writers draft the briefing. An education editor reads alongside them for any term a reader would need explained.
- Review runs five desks at once: style, fact-check, legal, ethics and engagement. A managing editor reads every verdict and writes a memo that quotes its evidence. Then the editor-in-chief decides.
- Extraction turns the briefing into structured events in four passes. The pass that analyses those events never sees the briefing’s text.
- A bias check looks for a lopsided panel of sources, and for perspectives that all sound the same.
- Publication links every event to its sources and to the people and organisations in it. A separate reviewer checks the image.
- Post-publication runs in a fresh session. It assigns the briefing to its stories, then updates and tidies the page of everyone it mentions.
A full run usually involves 20 to 25 department interactions. If a desk is skipped, the run cannot publish.
Checks
- Legal has the only absolute veto. A hold stops publication.
- Fact-check assumes every figure is wrong until it traces to a source. It hunts for invented quotes and suspiciously precise numbers.
- A style editor checks that the briefing reads naturally: varied sentences, plain words, no stock phrases. A script flags machine-like patterns to help it.
- The editor-in-chief never fixes a draft itself. It sends the failure back to the desk that owns it, often three or four times.
- A gate script runs about 20 named checks. Among them: every name in bold must match a known page, no event may retell one already reported, and every date must match its day of the week.
- New people are checked against outside reference data. Models like to invent a relative with the right surname and the wrong first name.
People, companies and places
Every person, company and place a briefing mentions gets a page, with a timeline of events and links to the others. The rules for those pages:
- Only real people, companies and places get a page. Statistics, abstract ideas and the outlets being cited never do.
- Each process that updates a page owns its own fields, and the code stops one overwriting another’s.
- A durable fact must be stored as evidence, or the next rewrite of the page loses it.
- A hygiene scan sorts defects into nine kinds, each with its own fix. Two kinds, suspected duplicates and changes to a person already verified, are escalated, never patched by an agent.
- A separate audit catches what the scan cannot see: a name with a page of its own that a briefing forgot to link.
Memory
- Story threads follow each story over time. Every run receives the recent events and the facts already covered, so a new briefing reports only what changed.
- Corrections never edit the original briefing. If a corrected claim reappears in a later draft, fact-check fails it.
- Every page keeps its history, so I can say what any page said on any date.
How big it is
| Part | Size |
|---|---|
| Departments | 21, with about 7,000 lines of instructions |
| Shared reference documents | 15 |
| Pipeline scripts | 72 Ruby scripts built on a library of about 69,000 lines |
| Named checks in the gate script | About 20 |
You can read the result at lowdown.today.