Prompt-first tools are good at a different job
ChatGPT, Claude and Gemini are genuinely excellent. Give any of them a well-formed question and you get a better answer than most people could write themselves. But they share one structural property: they are reactive. They do nothing until you open them, and they can only help with the thing you already thought to raise. The quality of what you get out is bounded by the quality of what you thought to put in.
That is fine when the thing you need is an answer. It is a poor fit for the failure mode most people actually have, which is not "I couldn't find the answer" but "I forgot the thing existed". Nobody opens a chat window to ask about the domain renewal they don't remember signing up for.
What replaces the box
Nova watches the sources you connect — email, calendar, WhatsApp — and does the noticing itself. Overnight it reads what arrived, sorts every item into one of five categories (Needs you, Waiting, Money, FYI, Noise), and for the things that need a response it prepares the response. By morning your home screen is a short list of cards, each with the next move already drafted.
So the input you give Nova is not a prompt. It's a decision: approve, edit, snooze, or dismiss. The clever part already happened while you were asleep. Your part is judgement, which is the part a model shouldn't be doing on your behalf anyway.
It prompts you because it learned you
There is a second half to this. A home screen that opens with cards instead of a box only works if the cards are the right ones — otherwise it is just a different kind of noise. So Nova learns from what you do with them. Editing a draft teaches it your tone; the kinds you consistently approve rank higher in the next morning's recompute; the ones you snooze or dismiss sink and eventually fold into the quiet section. It learns from the sources you connected and from your own actions in the app — nothing else — and every learned pattern is listed in Memory with a reset next to it, plus a full export. That is the honest bargain: Nova can prompt you instead of waiting to be prompted because it has learned which things are worth prompting you about.
The honest caveat: Ask exists
Nova does have a place to type. It's called Ask, and it works the way you'd expect — questions about your own connected data, drafting help, summarising a document you snapped. We didn't remove the text box from the product; we removed it from the home screen.
That distinction is the whole argument. What sits on a home screen is a statement about what the product thinks your default mode is. If the default is a prompt box, the product's position is that you should arrive knowing what you want. Nova's position is that you shouldn't have to. You can still ask Nova anything. The point is that you rarely have to.
What this design costs
- Nova is only as useful as the sources you connect. With nothing connected, a blank chat box would honestly be more useful than an empty card list — which is why setup leads with connecting one source.
- Nova can be wrong about what matters. Every card shows why it was surfaced, and correcting it teaches the ranking. But a prompt box is never wrong about relevance, because you supplied the relevance.
- Prompt-first tools answer instantly. Nova's core value arrives on a schedule — overnight, then through the day. It is an early-warning system, not a search bar.
Why we think this is where assistants go
Prompting is a skill. It is a learnable one, and plenty of people have learned it. But a skill requirement is a filter, and the people who most need an assistant — the ones already carrying too much — are the least likely to sit down and craft a good prompt at 6am. An assistant that only helps the people who know how to ask is an assistant for the already-organised.
So Nova inverts it. You don't prompt Nova. Nova prompts you — with something specific, already prepared, and one question: approve?