Tacit

Every AI notetaker on the market solves capture and calls it a product. Granola is a personal notepad. Otter is a sales tool. Fathom is a free recorder. All three are good at turning an hour of talking into a page of text, and none of them own what comes next, which is the part that actually hurts.

The pain was never the note-taking. It's that six weeks later nobody can remember whether procurement signed off, who owned the pricing brief, or which call the decision was made on. So the team re-litigates something it already settled. The transcript existed the whole time; it just wasn't findable by anyone who wasn't in the room.

Build the product around search instead of capture, and let that decision cascade.

The interface is a live notes-and-summary split, a searchable knowledge library, and an action-items panel that assigns commitments to the person who made them. The centrepiece is Ask Tacit: a plain-English question across every meeting the team has ever had, answered with the exact quote, the speaker, the meeting and the timestamp. A cited answer, not a summary, because the failure mode of a knowledge tool is confidently telling you something nobody said.

Two typefaces, ink and ember, chosen to stay calm while the screen is actively updating. And one commitment held across the whole thing: no bot in the participant list, no audio stored. It captures the way a person in the room does.

Both builds made with Lovable, kept as separate projects and held together by one token file rather than one codebase.

Keeping a screen quiet while it writes itself.

The transcript streams in live, the summary rebuilds underneath it, and neither is allowed to drop frames, reflow the paragraph you're halfway through, or steal your scroll position. That is a typography problem and a rendering problem arriving at the same moment: line height, measure and colour decide whether streaming text reads as calm or as flickering, and the render path decides whether it reads at all.

There is no handoff at which that gets solved. A designer specifying it without the render behaviour writes something that judders; an engineer building it without the type decisions ships something correct and unreadable. It is the clearest case on this site for one person holding both.

The second hard part was quieter: the site and the app are two separate codebases, on purpose. Marketing and product move at different speeds and shouldn't share a deploy. Brand continuity is enforced through shared design tokens instead of a shared repo, so crossing from the site to the app feels like one company, and neither build slows the other down.

A fictional product, designed and built to the point where the interface decisions are testable, including the mock world behind it, where the same commitment appears in the summary, the action items and the search result, with the same owner and the same due date. A demo that contradicts itself is just a mood board.