Fight the Line, Not the Thread

I was three paragraphs into a long Ask answer on my phone when I found the one line worth arguing with. The rest of the reply was scaffolding: context I already knew, a summary I did not need repeated. One sentence, buried in the middle, made a claim I wanted to push on. My thumb was already moving toward the reply box, ready to type something like the bit about the retry logic. Then I stopped. Why was I about to describe a sentence that was already sitting on the screen in front of me?

Highlight-to-quote is the fix I put live instead. Drag a selection over a passage in an Ask reply. The interface pins that exact span as the subject of your next message, shown as a small quote chip in the box. The model receives what I call a focus block: the excerpt, placed ahead of your follow-up. That is how it knows which sentence you mean without you retyping it. Fight the line, not the thread.

The Reply Is Treated as One Block Because That Is Easier to Build

Selection is a second cursor. Chat products default to a single unit of follow-up because a single unit is what a demo needs. One input field, one growing transcript, one send button: nothing to model beyond appending the next message to the same list. Building a second unit, a passage inside a reply that a reader can address directly, means answering three harder questions. What counts as an addressable span? How does it survive a re-render? What happens when the model producing the text has not finished the sentence yet? None of that shows up in a first demo. The bubble is enough to convince a room the product works.

That default rewards whoever releases first, not whoever answers best. A five-paragraph reply that buries one wrong number in the third paragraph gets the same follow-up box as a one-line reply. The reader question I opened with, how you ask about one paragraph without retyping it, has exactly one answer inside a product built this way. Type it again, and hope the model works out from your paraphrase which paragraph you meant. That is not laziness on the reader's part. The interface never gave them any other way to point. Prose is a slower, blunter pointer than a thumb already resting on the passage.

Messaging Apps Already Solved This. Chat Products Forgot.

There is a genuine case for sending everything. A model reasoning over the full conversation has more material to work with than a model reasoning over four words a reader dragged across a screen. More information should, in principle, produce a better answer. Set that against a pattern most readers already carry in their pocket. In June 2016, WhatsApp added quote-and-reply for one message inside a long group thread. Most people would not open their first AI chat window for another six years. Engadget covered the rollout as it left Android beta for the stable apps. Nobody using that gesture has to retype who or what they mean.

Anthropic and OpenAI ran a version of the same fix on the document side of a chat product. Artifacts opened a second panel beside the conversation in June 2024. It let a person point at a paragraph the model had just written instead of re-describing it in a prompt. That is a distinction I made in Familiarity Is a Floor, Not a Ceiling. OpenAI answered four months later with Canvas, built for the same reason. Both panels fixed this for a document sitting beside the conversation. Neither fixed it for the conversation itself. That is the one surface Search Finds Pages. Chat Finds Answers. named. That essay argued the surface should be judged on a cited answer, not a page of links. An answer worth citing ought to be an answer you can point back at.

The competitors I watch most closely already put the gesture back on the reply. Perplexity labels it Add to follow-up and parks the selection above Ask a follow-up. Claude drops the same span into the composer as a quote block before Write a message. Both treat the answer as the surface you point at, which is the unit Artifacts and Canvas left behind.

What the Quote Chip Cost to Build

Building the gesture meant making decisions nobody reads a product page for. The floating Ask about this button has to hit a 44-point touch target. The phone screen is about to be crowded by a keyboard. An excerpt has to be long enough to mean something and short enough not to become a second essay. The interface accepts a single word and rejects anything past 800 characters, with a toast that says which limit you hit. The highlighted span itself has to survive the interface redrawing around it while you keep reading. That meant reaching for a browser highlighting method most readers have never heard of. Browsers that do not support it yet need a fallback. None of this is a feature idea. It is the argument, written in code, that a passage deserves the same care as a whole reply.

The riskiest decision was the smallest one: a teach tip that shows once, in a single session. It tells a reader to select text in an answer to ask about it. I do not yet have the numbers on how many people see that tip and act on it. I am not going to invent a percentage to make this paragraph land better. I have already made this argument about other people's changelog modals: a good release still fails if nobody notices what changed. A one-line tip that fires once is exactly the kind of good idea a busy reader dismisses without reading. I do the same thing to every What's New modal I did not ask for. Somewhere on Ask this week, a reader is retyping the paragraph about the toast validation into the reply box, one clause at a time. Nobody showed them they could point at it instead.

You Point Before You Can Say What You Mean

Watch a child whose thumb is already moving before they have the word for what they want. They do not construct a sentence. They raise an arm and open a fist toward the shelf, then wait. The pointing works faster, and more precisely, than any word they will learn to say instead. Language arrives later, mostly as a patch for the times nobody is close enough to see where the finger is aiming.

A highlighted passage is that same gesture wearing a cursor. It gives the reader a kind of reference no amount of extra context in the model's memory can substitute for. A description is always a little late and a little wrong, built after the fact from memory of what you meant. A point is exact the moment you make it. Selection puts that gesture back into a reply built entirely out of language, the one place where pointing had quietly gone missing.

The Strongest Case for Sending Everything

The fairest version of the objection is this. Send the model the whole thread, every time, and never let a reader narrow what the model sees. Every prior turn gives the model more to reason with than four words dragged out of a paragraph. The sentence right after the one a reader highlights is sometimes the qualifying clause that made the highlighted sentence true in the first place. Excerpting looks like precision. It can just as easily be omission wearing precision's clothes. The reader doing the highlighting is the person least likely to notice what they left out, because they already believe they know which part matters.

That risk survives contact with the interface only if quoting means replacing. It does not. Ask sends the model the full conversation exactly as it stood before the selection. It adds the excerpt only as an instruction layered on top. Treat this passage as the subject; do not ask which article or topic the reader means. Nothing upstream of the quote is deleted, trimmed, or hidden from the model reasoning over the reply. Pointing at a passage tells the model where the reader's attention is. It was never a request to look away from everything else.

Fight the Line

Fight the line, not the thread. A quote chip did not add a feature to Ask; it gave the reader a second cursor for any product still charging the blank-box tax one paragraph later. Somewhere on Ask this week, a reader will drag four words across a sentence rather than retype them from the reply box.