The search box is such a familiar object that it's easy to forget how strange it is. To use one, you stop what you're doing, translate a half-formed thought into keywords, and context-switch into a results page. Decades of practice made us good at this, which hides the fact that the skill exists to compensate for the interface. The query is a compression format for curiosity, and like most compression, it's lossy.

What made me finally act on this was watching how people read AI answers. An answer almost always spawns the next question: a term you don't know, a claim you doubt, a product you want to see. And that's precisely where the thread dies, because following up means composing a new query. I kept seeing the same pattern in sessions. Curiosity arrives mid-sentence, and the interface makes it walk back to the box.

Selection as the search gesture

There's a useful way to lay out the options for moving through information. Hyperlinks are author-defined paths: you can only go where the writer anticipated. Query boxes are reader-defined but high-friction: you can go anywhere, if you pay the composition tax. The missing quadrant is reader-defined and near-zero friction.

Selection fills it. Pointing at a phrase is the lowest-cost gesture we have for saying "this, right here, is what I'm curious about," and unlike a typed query, it carries its full context along. The system sees not just the selected words but the sentence, the answer, and the conversation they sit inside. Ambiguity that would wreck a two-word query resolves itself when the query never leaves its context. A selected span is a better query than most typed queries, because nothing was lost in translation.

So I built selection into Komo as the primary follow-up gesture. Select any text in an answer — or click an entity the system has highlighted — and a small menu offers four verbs:

• Learn more — dive into the selection with the surrounding context carried along
• Quote — ask a follow-up question scoped to exactly this passage
• Websites — get direct links for the selection, for the times you want sites, not answers
• Explore — hear what actual people say about it

The four verbs map to the distinct intents I'd already observed in search behavior (understanding, interrogating, navigating, and opinion-seeking), so the menu is a miniature of the whole search engine, scoped to a phrase.

Design notes

Two decisions mattered more than I expected.

The first was ruthless sameness: the identical gesture works in quick answers, long research reports, and explore feeds. Interaction techniques earn their keep through habit, and habit needs consistency. Learned once, the gesture becomes ambient, and people stop noticing they're "searching" at all.

The second was restraint in the menu. Early sketches had more actions, and every addition made the common cases slower to find. Four verbs that cover the intent space beat nine that enumerate features. The discipline of cutting here felt like the interaction-design equivalent of editing prose.

The observation that pleased me most in testing: downstream results from selections were consistently sharper than results from typed follow-ups on the same topics. In hindsight this is what the design predicts, since the selection is the intent, unmediated. But it was satisfying to watch an interaction principle show up as a retrieval-quality effect. Better input affordances make better AI, without touching the model.

Where this points

I think of this project as one move in a larger argument: search shouldn't be a destination you visit. Reading generates the questions, so the page you're reading is the most natural place to ask them. Selection works inside an answer engine today, and the same logic applies to documents, code, and conversations. I'd like to see the gesture travel that far.