The five modes

The five modes were once the product’s headline. Since the restructure they’re something less visible and more useful: the response contract underneath. You never choose one. Recognition decides, shows you its choice as a chip on the capture card, and one tap flips it if it got you wrong.

They’re still worth writing down, because they’re the answer to a question every AI product has to answer: what shapes of response does this system have? Five, deliberately, a closed set, chosen once, defended since. Every request to add a sixth has been refused.

Before any of them, recognition works out what it’s looking at: a fast pass that decides the shape of the response, and a second that works out where the thought lives, what it connects to, and when it should come back. Both use provider tool-use with strict schemas validated by Zod, with one retry-and-repair attempt and a safe fallback, never “please return JSON” and hope.

Note: file this for later. The default fate of a thought: interpreted, filed into the wiki, and given a surfacing rule (resurface in a briefing, on a date, or in a context). About three quarters of captures end up here, which is the single most important number in the whole project’s history.

Brief: give me a fast answer. A streamed answer that shows clarification cards inline when the question is ambiguous, instead of guessing. You can critique the answer, get a revision, and accept the result into the wiki.

Think: reason through this with me. A longer conversation that re-grounds itself every few turns so it doesn’t drift, with bookmarks and branching, and at the end, the whole thread is distilled into a single synthesis artifact rather than left as a transcript.

Build: cascade decisions into a document. It gathers context first (the destination and what’s already there), does as much work as it can alone, then comes back once with a draft and a small batch of questions, compounding into a versioned spec, plan, or draft.

Explore: research this properly. Web research built up as progressively cited sections, checking your own knowledge base before the open web. Your annotation blocks are structurally protected, the AI cannot overwrite them, and results save into the wiki with merge, replace, or new-page strategies.

Since the agentic layer, a single capture can draw on more than one of these: recognition emits a short plan of typed steps, and the modes are the vocabulary those steps render in. The set didn’t grow. It just stopped being a menu.