The engine, stated plainly

Your topics are not keywords to us; they are standing editorial assignments. 'Payments infrastructure' and 'the Premier League' each get treated the way a beat editor would treat them: what does this listener already know, what angle would actually be new to them this morning?

The engine keeps this light on purpose. We store your topics, tone, and briefs, and nothing else. There is no shadow profile, no engagement optimization, no inferred interests you never asked for. The rundown you get is a function of what you told us, which means you can always see and edit everything that shapes it.

Each morning a language model plans your rundown, then writes each chapter as a spoken segment in your tone, following a radio-style guide: one idea per sentence, verbs early, written for a single pass of the ear.

The output is not prose we hope is right; it is a structured episode validated in code. Every chapter must declare its basis: material you added (labeled with its name), or the model's own knowledge (labeled with its recency). A chapter that frames general knowledge as this morning's news, or arrives without a basis, fails validation and never reaches you. On contested topics, the style contract requires attributed, balanced framing.

Today, briefs compose from your topics using the model's knowledge plus anything you paste in. Live news-source connectors, which feed the same pipeline with dated articles as source-labeled material, are the production integration and are documented in the README and roadmap. We say this plainly because a briefing product that is vague about its sources does not deserve your 7am.

The finished script is synthesized chapter by chapter through ElevenLabs, with settings tuned for early mornings: steady, warm, unhurried. Punctuation is honored the way an announcer honors script markings, which is why the writing stage writes for the ear in the first place.

Audio streams to the player per chapter, so skipping never wastes synthesis, and identical chapters are cached and never synthesized twice. If voice is unavailable, the readable brief stands in with an honest note; the product never ships a silent failure.