Chat Is the Future

In the summer of 2025, a new browser war was brewing. Perplexity launched Comet, and OpenAI followed that fall with ChatGPT Atlas. Both competed with Dia, but I was particularly impressed with Comet and Atlas because they could browse the web for me. I handed them a grocery list and asked them to fill up my Whole Foods shopping cart. Even though the process was slower than if I’d shopped manually, it was an aha moment. I wanted to be able to control my computer, the websites I visited, and the apps I used with plain language.

During my research for this essay, I went to Amazon to see if they’d integrated something like this natively on their site. And they did! Alexa for Shopping helps you shop on Amazon via conversation. I tried my grocery list experiment again. Within 15 seconds, my cart was filled with the dozen items on my list. Last year, Atlas took over 11 minutes and Comet took four.

There were two mistakes, both worthy of no more than a shrug. I’d asked for mild Italian sausage and Alexa found a package of frozen, which I wouldn’t have gotten. But to be fair, its options were slim. It also added two bunches of carrots instead of two loose carrots. Easy to correct in the cart.

The more I use apps in 2026, the more I want to start by telling them what I need. By chat, I mean expressing intent in plain language, typed or spoken. Give it the grocery list. Describe a design to make. Then use the controls when there’s something to inspect or change.

I’ve wanted those controls all along. What’s changed is how often I want to start with plain language.

The Long Case Against Chat

In January 2023, Austin Z. Henley wrote:

Expecting users to primarily interact with software in natural language is lazy.

Henley rightly argued that chat puts the onus on the user of figuring what to ask and how to ask it. Good interfaces guide people and help them discover what’s possible. He doubted that natural language should be the primary interface.

That same year, Amelia Wattenberger wrote:

Good tools make it clear how they should be used. And more importantly, how they should not be used. If we think about a good pair of gloves, it’s immediately obvious how we should use them. They’re hand-shaped! We put them on our hands. And the specific material tells us more: metal mesh gloves are for preventing physical harm, rubber gloves are for preventing chemical harm, and leather gloves are for looking cool on a motorcycle.

A blank text field doesn’t give the user a lot to go on. Wattenberger suggests that a chatbox with some additional affordances could help: dropdown menus for settings, or sliders for parameters to shape the response.

In January 2024, Elliot Smith predicted that chat would eventually occupy a narrower role:

Chat is unlikely to disappear but it’ll likely be for power users, much like the command line is today.

By 2025, Allen Pike was arguing that “chatting with the underlying LLM should be more of a debug interface—a fallback mode—and not the primary UX.” And at the time I agreed.

I’d already been arguing for a hybrid in my experiments with prompt-to-code tools. Starting with a prompt was useful, but once a design appeared, I wanted direct manipulation. At the time, apps like Lovable were very much prompting-only, unless you knew how to edit the code. As a designer, I wanted to click on some text and change its size.

I still agree with Wattenberger about the need for controls. But I’m having a harder time seeing plain language as a fallback or something mainly for power users. Pasting a grocery list into Amazon feels like a better way to shop. Just cut out the tedious middle part of searching, comparing, and adding to my cart. Just show me the outcomes. Increasingly, it’s the way I want to do all things on my computer.

Siri Finally Did Something

I’d given up on Siri save for timers and asking it to run my smart home scenes. Every morning, I walk into my home office and say, “Hey Siri, open office.” The lights and power strips switch on, which turn on my speakers and external display. At the end of the day, I say, “Hey Siri, leave office.” Siri enthusiastically answers, “On it!” The office powers down, and Siri then says, “Good morning!” Ugh. It’s so dumb.

So the other day, I decided to give the new Siri AI in macOS 27 a task. I had a list of a dozen contacts I wanted to add to my address book. I took a screenshot of the table containing names, emails, and phone numbers, pasted it into the new Siri AI app’s chatbox, and prompted, “can you make sure that all these people are in my contacts?”

It took less than 30 seconds. Siri read the text in the PNG, compared the names with people in my Contacts, and updated the information accordingly, adding new contacts as necessary. I loved this interaction.

Doing it manually would mean searching for each person, opening their record, comparing the information, and making the updates. If they weren’t there, I’d have to create a new contact and type everything in. Multiply that by 12 and it’s a good 6–10 minutes worth of clicking around.

Apple’s App Intents technology lets apps expose a menu of actions for the system to use. With the new Siri, I can describe what I want done and let it figure out the steps. I didn’t have to explain the loop for checking and updating each contact.

And there’s an actual chat app. I could attach the screenshot, which gave Siri the information it needed without my having to dictate a dozen names and email addresses. According to Apple, Siri can also use what’s on my screen to “get answers, search, or take action.”

Asking for something in plain language feels a lot more useful when the computer has the context and can actually do the work for me.

I Still Want the Controls

Being able to give Siri a screenshot made it easier to explain what I wanted. But what happens after the initial request is made?

When I wrote about generative UI, what interested me about Gemini was how it could answer a prompt with an interactive experience. Ask about the painter Vincent van Gogh and you could get a little website to explore. The AI chose how to present the answer, and I could interact with the result. The whole thing still started with me asking for what I wanted.

More recently, OpenAI introduced Intelligent UI with GPT-6, promising interactive responses and tools built on the fly. AI startup Hark combines proactive prompt suggestions with generative UI widgets. In an interview with Gizmodo, designer Abidur Chowdhury says:

Our goal as a company is almost to build a graphical user interface to this future technology—to build the mouse, which is like the interaction method with this future piece of technology.

Proactive prompt suggestions in Hark.

Hark’s suggestions are based on your personal context. Think of them as pre-made, custom, one-click prompts. They help with knowing what to ask. The widgets give you a more visual way to work with the response.

Bespoke widgets in Hark.

Will Hark be able to build a better mouse? Honestly, my money is on “no.” No shade intended. I just don’t think there’s a there there.

There are plenty of patterns, component galleries, and design recommendations for AI interfaces now. Showing progress, explaining tool calls, and presenting changes all help. But when I’m making something, I also need to edit the artifact itself.

In a recent essay, Wattenberger wrote:

As we build interfaces for these new AI models, we keep leaning on familiar ones. Increasingly, everything is chat. Much in the same way that everything used to be a command line.

We kept the command line as computing evolved, but we also added new ways to work (and think).

Chat is great for giving rough direction. It’s less suited to the fine-grained changes and little tweaks that follow. Sometimes I just want to reach in and move something.

“Chat is great for giving rough direction.” You have to start somewhere. I think it’s chat. Then, depending on what you’ve asked AI to do for you, whether it’s writing code, drafting a design, or generating an image, you want to have the tools to click and edit directly. “Sometimes I just want to reach in and move something.”

In my prompt-to-code tool shootout last year, I wrote:

We [designers] need to be able to directly manipulate the components by clicking and modifying shapes on the canvas or changing values in an inspector.

This year, things came a long way. When it launched, Claude Design’s primary interface was chat, but it critically had some affordances for direct edits. In my review of the tool, I noted:

I could select a piece of text on the page and directly change the copy or tweak the font, size, color, or any of the traditional properties of type. The changes happened immediately in front of me.

That’s important because making those edits directly is a lot easier and more precise than typing and describing the changes. Tweaks like these are great when you’re iterating and dialing in the details. But earlier in the process, the models have gotten good enough where I want to just let them cook.

Working Like a Creative Director

Recently, I built a new design system for Luma Health, a health technology company. The company was going through a brand refresh and wanted their product to share the same brand DNA. Naturally.

I put together a package to brief Claude Design, treating the tool like a junior art director. Here is the moodboard from the brand marketing designer, and here is additional creative direction from me. Now go make five directions, expressed as these two screens.

A lot of the first round came back as expected: generic, just color variations. So I refined the creative direction and asked for five more. The second round yielded two interesting directions, ones good enough to pursue further.

I iterated on those, but this time using the controls to tweak the background surface color or adjust the size of the type. Playing around with type size directly is a lot easier because I can quickly tweak and see if the new size feels right. Typing a sentence in the chatbox means waiting for the LLM to make the change before I can see it. With direct manipulation, the iteration loop collapses.

There were also changes broad enough to send me back to chat, like asking it to adjust the component color schemes to meet WCAG 2.2 contrast requirements. And out of exploring those two directions, a third idea began to emerge. I spun that off and presented all three to the client.

This way of designing fits with the last two decades of my career—being a creative director. I can brief the work, judge what comes back, and give it another round of direction. Then I can get into the details myself when that’s quicker. I want to move between those modes as often as the work requires.

Starting with the Outcome

Sometime last year, I played with n8n, the node-based workflow automation tool. I wanted to input an article URL and get a saved version of the article as a note in my Obsidian vault. Sounds simple enough, but I had to work out whether the link was a YouTube video, whether a generic scraper would work or I needed to use Apify, and how to create and save the Markdown file in Dropbox.

Honestly, it’s a pretty simple workflow as far as n8n workflows are concerned, but building out the node graph was tedious. Then I realized that I could use Claude Code to write the JSON that made up the workflow. While it did that, it also wrote custom Python to handle more complex cases like stripping out noise from the Markdown.

I’m currently designing a similar agentic workflow product, and I’m taking my experience as a user into it. A GUI for assembling these workflows means knowing which node type to use and how to connect it to other nodes. It means setting the right options and parameters in each one. And thinking through the logic like a programmer. Even for someone like me, who’s highly technical and logical—much to my wife’s chagrin—building these workflows is challenging. How can I ask a non-technical user to do that?

I want them to be able to describe the workflow they need. Then the app should draw it on a canvas, with controls to inspect and change it before it runs. The user still needs to understand what will happen. Asking AI to build the workflow doesn’t remove that responsibility.

Take my article-saving workflow. A suggested prompt could be as simple as “Save an article from a URL to my notes.” The resulting canvas could show where the content comes from, how it’s processed, and where the file will be saved. I could change the destination folder directly, or ask it to handle YouTube links too. I’d have something concrete to work with instead of a blank canvas and a menu of node types.

For the product I’m designing, we also need guardrails that keep the workflows within the platform’s capabilities and the requirements of the work they’re supporting. The interface has to make the consequential choices clear enough for someone to review them.

That still leaves the blank-page problem. How does someone know what to ask for? Suggestions, examples, and training still have a job to do. Hark’s contextual suggestions are one approach. In a product built for a particular kind of work, we can offer examples specific to that work.

The frontier LLMs have gotten a lot better at interpreting intent, so we don’t need to be as disciplined in how we write our prompts as we did a couple of years ago. Clear instructions and context still help, of course, but there’s less of a need to learn a special way of asking.

A suggested prompt can get someone to a sample workflow. From there, they can use chat and direct controls to tailor it to their own needs. That’s a much more inviting place to start than assembling the whole thing node by node.

Computer, Help Me with This

Sixty years ago, Gene Roddenberry’s Star Trek TV series showed the world what a plain-language UI could look like. In an episode about a traveling Shakespearean actor who might be a former murderous dictator on the run, science officer Spock asks for the personal dossiers on four people, then instructs:

Correlate. Check their past histories. Report any item, any past episode or experience they all have in common.

Sounds like an LLM prompt from today, doesn’t it?

As the series and movies continued, the computer became more capable. The crew could talk to it to investigate a problem or run a simulation. But the control panels stayed. They still used the helm to fly the ship and tactical controls to aim and fire weapons. Conversation and direct controls coexisted, and nobody seemed particularly troubled by that.

Montgomery “Scotty” Scott, chief engineer of the Enterprise, tries to use a “quaint” computer from the 20th century. (from Star Trek IV: The Voyage Home)

Of course, it’s all entertainment. But science fiction has a way of showing us what the future could be like. Asking the computer to do something in plain language was an ordinary part of working on the Enterprise.

In late September, OpenAI launched Dots, its always-on personal assistant product. Being able to type a request or talk it through on a call feels like another step toward that future. Here’s a demo by Dots product manager Holly Li.

As these tools get more capable, I expect to hand over more of the work by describing what I want. And I’ll keep wanting controls for the parts that are easier to do myself. Designers have plenty to figure out in the movement between those two modes.