LLenny's Podcast
← All frameworks
InnovationAparna Chennapragada

NLX: Designing the Natural-Language Interface

Conversation is an interface with real constructs — design it, don't just let the model lead

Difficulty
Moderate
Time to result
~weeks to results
Steps
4
Confidence
92%

'NLX is the new UX.' The mistake is assuming that with AI the model leads and you just chat, so nothing needs designing. Conversations have grammars, structures, and invisible UI elements. There is a concrete and growing set of natural-language constructs a product builder must deliberately design: the prompt, the editable plan, the shown work, and the proactive follow-up.

Origin

Coined and championed by Aparna Chennapragada (CPO, Microsoft); she calls it one of her 'Roman empires.'

Core principles

  • 01Elastic (natural-language) interfaces still require explicit design — elastic is not undesigned
  • 02A conversation has grammar, structure, and invisible UI elements
  • 03How much to reveal is a design decision, not a default
  • 04These constructs are a point-in-time tool and ripe for personalization

How to run it

  1. 1

    Design the prompt as a UI construct

    Treat the prompt itself as a first-class interface element, the way a dropdown or a menu once was — not as an afterthought.

  2. 2

    Return an editable plan for high-level goals

    When the user hands over a high-level goal, have the agent come back with a plan — preferably one the user can edit — as an explicit construct.

    Pro tip Editability turns a plan from output into an interaction surface for steering.

  3. 3

    Calibrate how much work you show

    Show progress / thinking-aloud, but tune verbosity: too verbose feels like watching a cron job's scripts; too terse and the user loses confidence the agent is on the right path.

    Pro tip Because inference is slow and the system is a black box, some visible reasoning builds confidence today — but plan to make it personalizable (TLDR vs full trace).

    Watch out Going fully silent during long compute makes users very uncomfortable.

  4. 4

    Design proactive follow-ups

    After a result, suggest the obvious next moves (e.g. after generating a black-and-white clip-art image, offer the natural refinements) to guide the user onto a happy path.

    Watch out Too many follow-ups annoys; too few loses the chance to direct the user — calibrate the amount.

In the wild

DeepSeek showing everything

DeepSeek went to the extreme of exposing its full reasoning trace, and users liked it — evidence that 'show the work' is a live, valued construct right now.

Validated visible reasoning as a design choice, while Chennapragada frames it as point-in-time and personalizable.

Image follow-ups

If a user asks to create an image and gets a rough black-and-white version, the interface should proactively suggest the obvious next refinements.

Turns a dead-end result into a guided path without over-prompting the user.

Common mistakes

'The model leads, so nothing is designed'

Assuming a chat product needs no interface design because the model just responds — ignoring the grammar and constructs of conversation.

Mis-tuning verbosity

Showing too much (feels like running scripts) or too little (user can't tell it's on the right path).

Is it for you?

Best for

Product designers and builders creating conversational or agentic AI products

Not ideal for

Deep changes inside mature GUI-driven software where a chat surface adds nothing

From the transcript

natural language interface NLX is the new UX

18:00

so conversations also have grammars they have structures they have UI elements they're invisible

18:30

prompt itself is a is a new construct and that's a new way that's a new UI element just like a drop-own was or a…

19:00

If it's too verbose, it feels like I'm running some crown job and scripts. Uh but if it's too tur, then I don't know if…

20:00

too much and you're kind of annoying me right like uh but too little and in in some sense you've lost an opportunity to direct…

21:00

From the episode

Microsoft CPO: If you aren’t prototyping with AI, you’re doing it wrong

Aparna Chennapragada