How it works
How a briefing gets made.
Briefly is not a model that reads the news to you. It is a pipeline with checks at each stage, and a rule that a claim without a source does not ship. Here is the whole of it, in order.
1. Reading
A registry of vetted feeds is swept every 30 minutes. Nothing is discovered by crawling the open web: a publisher is read only after it has been reviewed and added to that registry, and the list of feeds is public.
Everything read is embedded so that near-identical coverage can be recognised later — which is what stops the same announcement reaching you three times in three voices.
2. Grouping
Every couple of hours the day's items are clustered. Six outlets covering one launch collapse into one story with six sources attached, rather than six entries competing for the same two minutes of your morning. The cluster keeps all of its sources; you can open any of them.
3. Choosing
Once a day the clustered stories are ranked and a short lineup — four to six stories — is selected against your profile: your role, your industry, and the depth you asked for. Selection is the scarce resource in a briefing that is meant to end. Most of what happened does not make it, and that is the product working.
4. Writing, then checking
The lineup is written into a script — and then the script has to survive several gates before anyone records it:
- Every claim carries a citation A validator rejects the script if any claim has no source attached. This is a hard gate, not a warning.
- Two sources, or say whose claim it is A fact appears stated as fact only if it shows up in at least two independent sources. Otherwise it is attributed out loud — "per Anthropic's blog" — so you know whose claim you are hearing.
- A safety pass before anything is spoken A classifier reads the finished transcript for directive medical, legal or financial advice, leaked personal information, and instructions injected into a source trying to steer the hosts. A block sends the script back rather than recording it.
- Failure keeps yesterday live A rejected script is retried once with a targeted correction. A second failure routes it to a person — and yesterday's briefing stays up. Nothing unchecked is published to fill a slot.
5. Recording, and checking that too
The approved script is rendered by a licensed AI voice, then mastered into a single briefing. Synthesis can drop or invent a word, so the finished audio is transcribed back and compared against the script it came from. If the two have drifted apart, that section is re-rendered rather than shipped.
The voice is AI-generated and we say so — on this site, in the app, and out loud during onboarding. What is and isn't AI here is written out in full.
6. When you interrupt
Ask a question mid-briefing and the same rule applies to the answer. Relevant passages are retrieved first; if nothing retrieved is a good enough match, the honest answer comes back — that there isn't a strong source for it yet — instead of a fluent guess.
Answers are then checked back against the passages they cite. A claim that does not follow from its source is cut or the answer is generated again. You get the citation either way, so you can go and read the thing yourself.
When we get it wrong
These checks reduce errors. They do not eliminate them, and no one honest will tell you otherwise. When a briefing gets something wrong, the correction is carried in the next one rather than quietly edited away. If you spot something, tell us at hello@brieflylab.ai — a person reads it.
Hear it for yourself.
3 days free, then $7.99/week. Cancel anytime.