How to design AI interfaces beyond chat

1 The bottom line

Give your AI product one input that picks the right mode and model, and use context-aware elements instead of a blank chat box. Users do not want to choose tools or configure models, so every choice you remove lowers the load on them.

2 Key takeaways

• Replace modes and toggles with one input that decides what to do. Every picker pushes work from your product onto the user.
  • Josh Woodward, Google Labs25:40 says the lesson is to forget modes and toggles and use one prompt box.
  • Tara Seshan, OpenAI29:52 says Codex users miss nothing by not switching, and a billion ChatGPT users are being introduced to agents.
• Give the input context so the user never faces a blank box. A product that knows the user can answer without being prompted well.
  • Ethan Ethier, Built With Science46:30 describes Jeremy AI, a chat that also reads each user's workout, nutrition and progress data. One engineer built it in a weekend and it became one of customers' favorite features.
• Hide the machinery and name features as capabilities. Users care about the result, not the model, the tool call or the setup.
  • Roman Ugarte, Grok Bot27:35 says about 99% of automations are built through natural language.
  • Weber Wong, FLORA12:26 scrapped an alpha with a node for every model and rebuilt it with just text, image and video.
• Build shared primitives underneath so the surface can stay simple. Stackable blocks let the model pull pieces together without new screens.
  • Ami Vora and Mike Krieger, Anthropic20:31 say very few products get this right, and that when they do, agents start piecing together things that were not connected before.
  • Ian Silber, OpenAI31:00 lists systems thinking as a growing hiring criterion.
• Serve most people a simple default and keep advanced features for the few. One product has to fit very different users.
  • Ian Silber, OpenAI39:40 notes ChatGPT serves free users and people paying hundreds of dollars.

you kind of need a PhD in model pickers.

Tibo Sottiaux, OpenAI36:40

2.1 Cut the choices first

Modes, toggles and pickers move work from your product to the user, so remove them before you add anything.

• Let one input choose the mode and the model.
  • Set the north star as: the user types a task and the system picks the right harness and model Tara Seshan, OpenAI29:52.
  • Build one input that knows whether to answer fast, chat or go off and work Ian Silber, OpenAI48:00. Users should not have to think about models or speed.
  • Chat, ChatGPT work mode and Codex differ mostly in UI. Work mode runs on Codex underneath with the coding UI removed Tara Seshan, OpenAI29:52.
  • OpenAI is merging the chat and work toggle in ChatGPT. Feedback said people liked the work capability but sometimes wanted the speed of chat Tibo Sottiaux, OpenAI36:40.
• Remove model pickers and configuration from the product.
  • Follow Dots, which has no model picker. The only thing to configure is which channels to talk through Tibo Sottiaux, OpenAI36:40.
  • Tibo says even he gets fatigued by model pickers, reasoning efforts and multi-agent options Tibo Sottiaux, OpenAI36:40.
• Do not hand users a pile of tools to choose from.
  • Build primitives so the system knows the user and the UI feels accessible all the time, then iterate on how people react, say Ami Vora and Mike Krieger, Anthropic25:52.
  • Mike warns it is easy to transfer a bunch of cognitive load to the user Ami Vora and Mike Krieger, Anthropic25:52.
• Push your team toward simpler builds.
  • Inspire teams to build less complex things, because they keep striving toward complexity Tibo Sottiaux, OpenAI30:10.
  • Tibo counts this as something he got wrong. He sees simplicity as the direction of products, where the app nearly disappears Tibo Sottiaux, OpenAI30:10.

2.2 Replace the blank box with context

A product that knows the user and the moment can offer the right element instead of asking for a perfect prompt.

• Feed the product the user's situation so nobody explains it twice.
  • Build context and close the loop. Users will expect products to understand their situation without explaining it 27 times Leon Sasson, Rise46:16.
  • Train the AI on your own content and on each user's data, as Ethan Ethier, Built With Science46:30 did with Jeremy's videos and blog history.
• Show interface elements that fit what the user is doing.
  • Ship tap-through follow-up questions for open-ended asks Ian Silber, OpenAI44:40. They help users express intent without becoming prompt engineers.
  • Hold an email or letter in a writing block that the user can edit directly Ian Silber, OpenAI44:40.
  • Change the interface based on who the user is and what they are doing. Image generation gets its own tools too Ian Silber, OpenAI44:40.
• Let the product start the conversation.
  • Expect products to use themselves to give a gentle on-ramp, like self-driving Nan Yu, OpenAI28:00. A blank input box makes users ask what to do now, and then learn the product.
  • Add proactive suggestions from calendar and Slack Ian Silber, OpenAI48:00.
  • Let apps message the user and propose changes. Eugenia Kuyda, Wabi43:10 gives the example of new kids' events the app found.
  • Nan says this applies to general tools like ChatGPT and industry-specific ones alike Nan Yu, OpenAI28:00.
  • Eugenia says a consumer social product should feel alive, with lots of activity when you arrive Eugenia Kuyda, Wabi43:10.
• Handle settings through conversation instead of a settings screen.
  • Have the tutor ask "Do you want to set a reminder? I can set this for you," as Adam Turaev, Praktika54:50 does.
  • Adam sees this as how products will work in future, with agents handling more of the experience Adam Turaev, Praktika54:50.
• Consider voice where speaking is easier than typing.
  • Use voice because models are getting good and fast at it Tara Seshan, OpenAI27:14.
  • Voice changed the game when onboarding OpenAI's own people team onto a new product, and it cut the tech support Tara does for family Tara Seshan, OpenAI27:14.

2.3 Hide the machinery

Users want the work done, so keep the agent's internals, power-user tools and technology out of sight.

• Show progress, not internals.
  • Use a typing indicator and progressive updates, not tool calls, chain of thought or every click Roman Ugarte, Grok Bot14:50. You would not ask a teammate for second-by-second updates.
  • Early feedback asked for a visible to-do list, but nobody wanted the long text stream. Onboarding sessions confirmed it Roman Ugarte, Grok Bot14:50.
• Hide power-user abstractions like skills and slash commands.
  • Build skills as background primitives that bots create and use, so users never type a slash command Roman Ugarte, Grok Bot37:50.
  • Remove the hacky setup of home VPNs and Mac minis. It would not scale to millions of users or to businesses Roman Ugarte, Grok Bot37:50.
  • Claire Vo reportedly shut down all her OpenClaw bots and switched to Grok Bot Roman Ugarte, Grok Bot37:50.
• Run every agent in the cloud.
  • Decide early that the runtime lives in the cloud as a persistent colleague with its own computer and the same state everywhere Roman Ugarte, Grok Bot30:30.
  • Let users kick off work from a phone or text the bot, with no laptop kept awake Roman Ugarte, Grok Bot30:30.
  • Products that straddle local and cloud leave users with daily paper cuts Roman Ugarte, Grok Bot30:30.
• Pick a specific use case and hide the technology.
  • Avoid labeling products as AI and avoid pushing the newest model, because most people outside tech dislike AI and do not know models Dris, Methods46:00.
  • Plaud sells a $130 recorder for professional sales calls with a $20 a month app, and reportedly made about 250 mil ARR in 2024 Dris, Methods46:00.

2.4 Build shared primitives underneath

A simple surface stays simple when everything under it is built from the same few blocks.

• Route every action through shared plumbing so agents can do what humans can.
  • Use the same plumbing for agents and your REST API, say Ami Vora and Mike Krieger, Anthropic20:31.
  • Test other products by asking them to do something that takes real thought Ami Vora and Mike Krieger, Anthropic20:31.
  • Once the primitives exist, you can evolve from a side panel to a malleable personal landing page without new infrastructure each time Ami Vora and Mike Krieger, Anthropic20:31.
• Build composable blocks and reuse what already exists.
  • Design building blocks that stack on each other, as Notion did early on Ian Silber, OpenAI31:00. It keeps the product feeling like one simple thing.
  • Tell designers to do less. Check for an existing component, or extend an existing feature, before designing something new Ian Silber, OpenAI32:35.
  • Integrate with specialist tools instead of rebuilding them. Codex built itself an extension for Premiere Pro when a videographer asked it to edit video there Andrew Ambrosino, OpenAI58:00.
  • Codex's in-house videographer Brent edited all the Codex app launch videos this way Andrew Ambrosino, OpenAI58:00.
• Name features as capabilities, not buttons.
  • Ask what each project's launch post would say. If it is not compelling, reconsider it Roman Ugarte, Grok Bot27:35.
  • Write "Grok Bot can now," not "Grok Bot now has." It pushes the team to kill pixels Roman Ugarte, Grok Bot27:35.
  • Automations are defined in natural language, not through a trigger-and-action sidebar Roman Ugarte, Grok Bot27:35.
• Use long-lived named agents, not one-off chat threads.
  • Group work by role or swim lane, as on a team, instead of a new chat per task Roman Ugarte, Grok Bot49:30.
  • Copying and pasting between chats is a sign the model is wrong Roman Ugarte, Grok Bot49:30.

2.5 Fit the interface to each user

Give most people the simplest path, and spend your rebuild effort where every user already goes.

• Serve an extremely simple default and keep advanced features reachable.
  • Stream the simple experience to most people and keep cutting-edge features accessible for those who care Ian Silber, OpenAI39:40. Usage ranges from recipe ideas to automating a farm.
  • Test new things early with the power audience, such as the desktop app and Codex, before widening them Ian Silber, OpenAI39:40.
  • Cowork is now stepping into a broader audience Ian Silber, OpenAI39:40.
• Rebuild the feature every user touches first.
  • Check usage data. At beehiiv, 20% of users use automations and 10% have a podcast, but 100% use the editor Tyler Denk, beehiiv44:50.
  • Rebuild it with AI, styling and customization, and add Copilot across the app for admin work like analysis and sponsorship outreach Tyler Denk, beehiiv44:50.
• Point AI at the user's admin burden, not at the creation they love.
  • Let the AI take analyst work, sponsorship outreach and manual tasks. Copilot has full context of the user's business and content Tyler Denk, beehiiv46:10.
  • A creator on a panel said she spent two days creating and five on admin. Tyler wants five days creating and two on admin Tyler Denk, beehiiv46:10.

2.6 Make it feel human and capable

The last layer is how the product feels: like a teammate, with a face where fear is the barrier, and making users look good.

• Settle product debates by asking what you would want from a human teammate.
  • When both sides sound sensible, zoom out and ask how a colleague would behave Roman Ugarte, Grok Bot40:00. The answer is usually near unanimous, and then the team builds it.
  • A huddle-style voice experience with screen sharing, then async follow-up, came from this question. The team calls it being "colleague-pilled" Roman Ugarte, Grok Bot40:00.
• Use a face to lower anxiety, not as a gimmick.
  • Build avatars that simulate being watched, because fear of speaking to another human is a big part of the language barrier Adam Turaev, Praktika30:51.
  • Avoid perfect realism. A Pixar-style look sidesteps the uncanny valley Adam Turaev, Praktika30:51.
  • In onboarding, the fear option was one of the most clicked. Even the 2022 version, with pre-recorded Synthesia avatars and hand-built conversation trees, had an avatar reacting to users Adam Turaev, Praktika30:51.
• Frame AI features as making users look good, not judging them.
  • Let users publish the work of themselves plus the AI and earn bragging rights, as Claude Code and Codex do Max Schoening, Notion47:40.
  • Automatic code review tools that roast your pushed code feel bad. Kubernetes beat Heroku despite a worse user experience by making ops teams superheroes Max Schoening, Notion47:40.

2.7 Where they disagree

The tips split on how much to lean on voice and chat, how much to hide, and what AI should be for.

• Voice first, or visible UI first.
  • Flip the order so voice is primary and touch is the backup, once voice has real intelligence and memory Tony Fadell, Build Collective1:07:45.
  • Keep the UI, because most people do not want to chat when they just want to solve a problem quickly Leon Sasson, Rise46:16.
  • Tony expects near-term devices to have displays, because people need to see maps and visual information and do not yet trust AI enough Tony Fadell, Build Collective1:07:45.
  • Chat suits the highest-agency users, and the average consumer's ideal interface is closer to TikTok Anish Acharya, a16z35:15. Brian Chesky started a lab on next-gen interfaces.
• Hide the machine's steps, but show the user's own process and taste.
  • Design around the editing process, not the first draft Edward Tian, GPTZero53:10. GPTZero is building writing replay, authorship history and chat-history grading.
  • Have an LLM describe a region of the user's embedding map in words, then let users add or remove topics, as Adam Mosseri, Instagram35:15 does.
  • Return several expert perspectives with attribution and a summary, so users make up their own mind Casey Winters, SuperMe13:25.
  • Instagram's next steps include non-topical controls like more fun content or fewer photos in a row Adam Mosseri, Instagram35:15.
• Save users time, or help them spend it well.
  • Aim consumer AI at feelings, not productivity, because more people want to spend time than save time Anish Acharya, a16z32:30.
  • Make the app feel like it is working. Users stick with apps that make them feel performative and effective Dris, Methods34:25.
  • Anish cites a device that tracks how often a family laughs, and a screen time system that rewarded his son for saying "I love you" Anish Acharya, a16z32:30.
  • Duolingo has hundreds of millions of users because people feel like they are learning, in Dris's view despite not teaching languages well. A single-purpose wrapper like a calorie tracker feels easier and more trustworthy than prompting ChatGPT Dris, Methods34:25.
  • Tibo says the promise of AI is to cut noise so people spend attention where they want, not to add one more prompt per second Tibo Sottiaux, OpenAI19:58.