Every question an assistant asks spends a limited resource: your willingness to be interrupted. Ask too rarely and ambiguous notes fossilize; ask too often and the user turns the feature off. The design question is not how to ask—it is when a question earns its cost.
When a question earns its place
The strong cases share two properties: the ambiguity will matter later, and only the author can resolve it.
- An unresolvable referent in a note with likely future use: ‘ask him’—three candidates exist.
- A commitment missing its owner or date: a promise that cannot safely become a task.
- A contradiction with earlier notes: today’s decision reverses last month’s—which stands?
- Media without purpose: a saved photo or link whose reason will be gone in a week.
When silence is correct
Most notes deserve no question at all. Asking about them converts a memory tool into a chore.
- The answer is inferable from existing notes or obvious context.
- The note is low-stakes: a passing thought, a grocery item.
- The user is mid-capture—never interrupt the act of writing or recording.
- The same note has already been asked about once. One question per note.
Never during, rarely after
Timing matters as much as selection. The right moment is a quiet one, soon after capture but never inside it—a gentle queue reviewed when the user chooses, not a popup that hijacks the moment a thought was being saved.
The user trains the threshold
Every dismissed question is feedback that the bar was too low; every answered one confirms it was right. An assistant that does not learn from ignored questions will eventually be ignored entirely.
The one number that tells you if the threshold is right
Answer rate is the metric that matters, and it should be high. If most questions asked get answered, the system is selecting well. If most get dismissed, it is not asking too little—it is asking about the wrong notes, and the dismissals are training you to ignore the queue rather than teaching the system anything.
A useful sanity check on volume: over a normal week of capture, the number of notes that genuinely warrant a question is small—typically a handful, not a daily prompt. Anything approaching one question per note has stopped being clarification and become data entry with a friendlier name.
Silence at the wrong moment has a cost too, and it is easy to miss because it is invisible. A commitment left without an owner never becomes a task, and nothing signals that. The asymmetry argues for asking about commitments and decisions even when confidence is moderate, and asking about almost nothing else.
Ask only what the future will need, only what the author alone knows, only once, and never mid-thought. Every question that fails this test is friction wearing a helpful mask.