Modern LLMs have their roots in translation — taking one set of symbols and producing another that preserve the intent of the original. Having ingested essentially all symbolic content across the internet (and much besides) — they seem able to translate between all sorts of languages. Natural languages, of course, but crucially, also machine languages.

In this post, I will argue the following:

1. A significant portion of UX design is about translating intent.
2. LLMs are a powerful tool for translating intent.
3. Thus, translation capability represents a potent new tool in the UX design toolbox.

But first, an anecdote.

Translate to Javascript Please

A few weeks ago, my daughters got their hands on my keyboard, and noticed that there are letters (!!). The idea for a little web toy that displayed letters upon key pressed popped in my head. Then I thought, wait, Claude can probably oneshot this.

I want to make a web page that’s toy for toddlers learning to recognize letters. All the page has to do is show one letter at a time - the latest letter that was pressed on the keyboard, shown in upper case.

Other requirements:

• Make the letter always white,
• Change the background to some random pastel color on each key press
• Rotate through several different type faces. Pick some common ones from Google Fonts
• Make the letter about 80% as tall as the screen.
• Make the page response to change in the screen size

Here’s what it produced.

Having already used Claude + Github Co-pilot a bunch for coding tasks, I was not surprised, but still delighted, especially by how little special encoding was required to coax this out of the LLM. Claude took my intent that I expressed in English, translated/expanded it into HTML/CSS/JavaScript (a specification in a programming language), which is then interpreted by the browser as a webpage. The girls got a new toy!

LLMs as a Design Opportunity

I’ve spent most of my career designing tools for specialists — data analysts, data scientists, machine learning engineers. In my experience, translating intent is more than half of the UI’s job. Viewed through Don Norman’s Seven Stages of Action framework, translating intent is at the core of the Gulf of Execution.

• Gulf of Execution
  • Form the goal
  • Form the intention
  • Specify an action
  • Execute the action
• Gulf of Evaluation
  • Perceive the state of the world
  • Interpret the state
  • Evaluate the outcome

As the domain becomes more technical, the Gulf of Execution grows. Consider the following examples:

	User Intent	Input Method	Effort
Ride Hailing	"I am here. I want to go there."	GPS, Address Bar + Map	Not terrible.
Photo Editing	"I want to modify this photo in this particular way."	Toolbars of icons representing various effects. Menus upon menus of more effects. Layers for organization. Mouse, Keyboard.	Quite a bit. I need to learn what the effects are, and how they might be combined. Then navigate the interface to apply them.
Writing Software	"I want a piece of software that does X, given Y input."	At least one programming language. Writing text files, in a IDE (integrated development environment)	Yikes. I need a mental model of how the language is interpreted, and then to translate my idea into the language's syntax.

In all of these cases, I must learn/use a specific set of actions/symbols. The specificity required in the three examples grows progressively. So is the effort required, both in learning the abstractions required, and in the act of translating intent into the specific set of actions/symbols.

The last example is the most extreme. It is all translation work, and it also points to where I believe the opportunities are.

Design Opportunities Through the Lens of Translation

So where are the opportunities? My take:

• Domains where there is some specification/language that mediates the work, and…
• where there are lots of existing examples of the specification, and…
• where learning the specification, and generating it takes significant user effort, and…
• where verification can be (somewhat) automated, or intuitive

So, software, but also things like CAD, graphic design, data analysis, music and perhaps even architecture. To me, this is the exciting stuff. The barrier to entry on making is about to get a whole lot lower.

Verification UX will be a Differentiator

I left out legal as a domain, even though it fits will the first criterion re: language-mediated work, because it fails the last criterion around verification. LLMs are prone to hallucinations. Their output needs verification, and there’s no simple way to validate the output of LLMs to be correct in the legal domain.

In contrast, we already see LLM-based software engineering interface gain traction because verification can be partially automated. There’s the baseline verification that code runs. There’s also an existing ecosystem of tools/practise around testing and formal software verification. It’s not perfect, but it is enough that:

1. Users have clear methods of checking that the generated code works.
2. These systems can be trained and improved in a closed loop.

It is not hard to imagine other domains of technical design/specification, where there is some definition of “functional”, adopting this UX paradigm. PCB (printed circuit board) design and CAD for manufacturing comes to mind.

So, in this new paradigm, the UX of evaluating the generated design becomes the critical bottleneck. The user’s job is not done when the specification is generated. The user must have confidence that the output is fit for purpose. As using LLMs to generate specifications gets easier, more and more will be generated. Imagine having 10x more stuff to review. New UX for verification must be developed, or we will drown.

Generate, Judge, Select, Repeat

For more domains with an experiential output like UI design, 3D modeling, or music, we (the humans-in-the-loop) are the judge of fitness for purpose. Thus the interface must let the user experience the output quickly, and flexibly. I expect that the chat-centric LLM UIs of today will give way to more experience-centric UIs. The natural language interface to the LLM will move aside, perhaps even become invisible if/when voice control becomes viable. I suspect the core interaction loop will start to feel like this:

1. Prompt for generate
2. Review options and make a Judgement
3. Select a direction
4. Repeat

The UX design adventure of the next half-decade will be seeing where natural language falls short, and how LLMs based interactions will live along side direct manipulation (a la Photoshop/Figma/Final Cut Pro). One limitation I already feel is the friction in calling the LLM’s attention to specific objects/aspects of a design in a chat message. Think about how we can drag to select a bunch of elements in Figma today — using the same gesture to provide context to the LLM cannot be far behind, can it?

Thinking Conversationally

Being a student of Paul Pangaro, my mental model of interaction design is a conversation between a human and a system. UX design is about facilitating that conversation in service of the human.

In the WIMP paradigm, humans bore the burden of translating their intent. With LLMs, interfaces can begin to take on part of that burden. While I don’t anticipate LLM-based UX becoming dominant, they will be transformative in domains where “translating-to-machine-specifications” is a critical barrier.

In this conversation/cybernetics frame, this essay summarizes to:

1. LLMs lowers (potentially dramatically) the human effort required to form messages that the system can understand. (Or, the effort between Form the Intention and Specify an Action in the Norman framework.)
2. In specification-mediated work, LLMs will make generating new work much more efficient
3. Efficient generation of new work will increase the volume of work produced, putting pressure on the Gulf of Evaluation part of the loop.
4. Differentiation will come from complementing LLM-based translation/generation with UX focused on evaluation and review, i.e. how the system speaks back to us.

I wince at the thought of a world where review and verification is the majority of work. I suppose that’s an interesting design challenge as well. Nonetheless, an exciting times to be designing.

---

Inspirations

1. What is Interaction? - Dubberly, Haque, Pangaro
2. Prediction Machines - Agrawal, Goldfarb, Gans
3. “Everything I built with Claude Artifacts this week” - Simon Willison… is just an incredibly productive person