Conversational interface

Start with the fence, because two words are doing different jobs. A natural language interface is an input MODE: you ask for what you want in words, and it can live anywhere, including inside an entirely ordinary screen full of buttons. A conversational interface is a SHAPE: the transcript is the screen, turns are the only unit of interaction, and there is nothing else to press. Every chat assistant is both. A search field that understands a sentence is only the first.

The shape’s defining problem is that it has no affordances. A screen made of controls advertises what it can do just by existing, and a text box advertises nothing, so a reader looking at a blinking cursor has to guess at both the capability and the phrasing. Everything this pattern has learned about discoverability is an answer to that. Starter prompts in the empty state, suggestion chips under the last reply, an opening turn that lists what is possible in three concrete examples rather than in a claim about being helpful. Those are not decoration around the conversation, they are its menu bar, and a build that ships without them is asking every reader to invent the interface for themselves.

The transcript is doing more work than it looks like. It is the state, so there is no settings panel to consult: what is true is whatever was last agreed, in words, some distance up the scroll. It is the undo history, since correcting a turn means saying so rather than pressing anything. It is the only record of the agreement, which matters enormously when the conversation commits to something (a booking, a refund, a piece of code), and it is the navigation, because scrolling back is how a reader returns to a decision. Design it as a document, then: quotable, scannable, with each turn’s own boundaries clear. The parts have their own words, and the site holds them already: a turn is a chat bubble, the pause before a reply is a typing indicator, and the box is a prompt input.

The failure NN/g keeps measuring is worth memorising because it does not look like a failure while you are building. A conversation is linear, so it works beautifully for the path the team walked through, and collapses the moment the reader steps out of it: a question asked two turns early, a change of mind, a request that needs three fields at once. Comparing five options in prose is worse than a table. Picking a date in prose is worse than a calendar. The strongest conversational interfaces are therefore not purely conversational at all, they hand you a control at the point where a control is better, and the judgement about where that point falls is most of the design work.