In this talk, the speaker reflects on how AI is reshaping everyday software design, starting with a critique of the original ChatGPT interface and how its poor UI broke established messaging conventions. He argues that successful AI products must respect familiar design patterns while introducing new, intuitive ways to fine‑tune AI outputs. The discussion moves to AI‑generated imagery, highlighting the cumbersome Discord workflow of Midjourney and showcasing Krea AI’s visual fine‑tuning tools (style selection, brush‑based edits, and real‑time adjustments) as a model for future AI workflows. Next, the speaker introduces Magic Patterns, a prototyping tool that generates real, responsive code from prompts and lets designers visually fine‑tune UI layers, bridging the gap between design and code. He then details his own project for Hatch—a chatbot builder that replaces prompt‑heavy configuration with familiar UI controls: tone sliders, knowledge tabs for FAQs and documents, and a decision‑tree workflow that encodes business logic (e.g., routing to sales reps). Throughout, he emphasizes a design philosophy that favors familiar, click‑or‑drag fine‑tuning over complex prompt engineering, while acknowledging that prompt engineering will still have a role. The talk concludes with calls to action (follow on Twitter, check out Numi.tech) and a light‑hearted sign‑off.
Chapters
AI-powered software. It's everywhere now, but it's not just ChatGPT anymore. It's quietly transforming the tools we use every day, and I'm noticing new and interesting design patterns that have completely made me rethink how I'm designing user interfaces. But what emerging design patterns do I think will win, and how did I personally go about designing an AI chatbot Builder. When I first learned about ChatGPT, I was so excited to use it, but then I saw their interface.
Oh my God, what is that?
Oh my God, what is that? And I was quickly disappointed because it looked like an old piece of software that hadn't been updated since 1998. And I wasn't just talking about the aesthetics. It broke some obvious design patterns. Usually messages are surrounded by bubble-like containers, but instead ChatGPT had this weird off-colored row. Usually my sent messages appear on the right and my received messages appear on the left. But ChatGPT aligned all messages to the left. What was up with that? And this icon, why is it a paper airplane? Am I sending an email or a message? So I actually decided to redesign their interface to address some of these issues as a fun side project. But the important part is that eventually OpenAI learned, redesigned their interface to follow commonly understood design patterns. They also got rid of their ugly green branding.
Bravo.
These changes made their interface feel more familiar with other messaging tools. But why does this matter? And what can this teach us about designing software overall? The takeaway for me was that creating innovative software doesn't always mean you're reinventing how a user interface should work. Sometimes it means you're actually following commonly understood design patterns that have been established for years, especially for interactions that everyone is already used to. Like sending messages. Okay, so this one kind of seems obvious, but what about the new and emerging design patterns? The ones that were born because of AI? When AI finally made it possible to make images with words, I was shook. And at the time, everyone was hyped about Midjourney, an AI image generation tool. So I decided to give it a try. But when I actually started using it, I had to go to Discord. Ew. Click on newbie channel. And then type in /imagine and enter a prompt? Like, what, what was up with that? But when I hit enter and I saw what happened next, it was weirdly magical, exciting, satisfying, but also disappointing. Because like, why was this so complicated? Why am I using Discord? Why does this experience suck? The thing I do remember was that I was really impressed with this. You could select your favorite image generated variants and fine-tune your desired result. And as time passed, I started to notice AI image generation tools create more ways of fine-tuning images. This is a new and emerging design pattern, which brings us to a tool that I think is doing fine-tuning best: Krea AI. I was introduced to this tool after watching Andy do a banger YouTube tutorial on how he uses Krea.
Uh, I really wanted these to be in a certain style. I could type that into the prompt here, but I can also just click on the styles. But you know, I was really hoping for like an Adventure Time style, so I can select that. And now I have a completely different, completely different vibe. I didn't have to prompt anything.
Did you, did you catch that? You don't have to finesse with the prompt, you just click on the style that you want.
We are going to just add a little birthday party hat on this dude because he is going to a birthday party. With Krea, you don't have to reprompt the entire image like with Midjourney, if you were doing this, you would have to draw that area and then you would have to say a crow riding a motorcycle wearing a party hat. Yeah, that's fine if it's a short prompt, but if it's something more complex, comes a little bit of a pain.
But one of my favorite examples of Krea's fine-tuning design patterns is this.
And so yeah, like, it'll give us some party hats, but what's cool about real time is that you can grab your paintbrush and you're like, you know, make the paintbrush a little smaller and give him a, a blue party hat. Then you can select both of those and drag them around and get different results.
This entire video is packed with gems. Go ahead and check out his channel after this video and subscribe. It's— he's awesome. See, I know what you're probably thinking: fine-tuning images, duh, that's obvious. But I actually think that this design pattern will become standard across all sorts of AI workflows. Workflows, like fine-tuning AI-generated code visually. This is Magic Patterns, a next-generation prototyping tool powered by AI. Yes, you could totally generate a user interface with a simple prompt, but that's just the tip of the iceberg. The thing that I'm most impressed about is, one, its ability to generate real working code from simple prompts, and two, its ability to fine-tune designs visually by selecting on a specific layer and simply telling the AI what you want to change. Then the AI updates your designs and updates your code automatically, which I think is super cool. So you can generate prototypes using HTML or use one of my favorite component libraries, ChatCN. And because it's real code that's being generated, it's fully responsive. And you could also export your prototype directly directly into Figma and then fine-tune your designs even further. Tools like Magic Patterns make me so excited about the future because designers everywhere can now access the power of code using AI. I've always felt limited by what I can create because I didn't know how to code, and now AI is changing that. So check out Magic Patterns and start experimenting with some of these new AI tools that genuinely feel like superpowers. So let's talk about one of the most interesting projects I've ever worked on. This is Hatch, a messaging platform that helps businesses automate and scale their sales pipeline. They're also backed by Y Combinator, which incubated companies like Airbnb, Coinbase, DoorDash, Instacart, Twitch, and so many other cool companies. Their founder, Chris, reached out to me and asked if I could help them design their AI chatbot builder where businesses would build chatbots that would respond to customer messages, calls, and marketing campaigns, all with the goal goal of helping them scale their sales. So I thought like maybe they're on the cusp of something big, something that would change everything. And this is what I had to work with: a poorly designed interface with complicated controls and just an unintuitive workflow. So I started by researching and gathering inspiration from other AI tools like ChatGPT, obviously, but also some new finds I found by using teardowns.ai created by Designer Tom, which has a library of all design patterns of different AI tools. So in my research, what I discovered was that no one has really figured out how to fine-tune AI chatbots. It's usually just like a prompt-heavy workflow with complicated ways to fine-tune the chatbot responses. And at Hatch, we doubled down on more clear and familiar ways to fine-tune, like just selecting a tone or creating a new one, sliding this control to dictate response time. And when businesses click on this knowledge tab, they could add frequently asked questions, upload relevant documents, and even link customer data to give the AI more context on who they're talking to. And they can simply test their AI chatbot right here. But the most fun and experimental design pattern that we created was this decision tree workflow where you can outline actual logic that the AI should follow in each conversation. Logic that could help the AI determine whether this conversation when it should be transferred to an actual human sales rep or ended because the customer isn't the right fit. Logic that could help the AI determine a set of actions like scheduling appointments, changing the conversation status, updating customer information, and so much more. What I realized was that this new and emerging design pattern of fine-tuning was actually not entirely new. It often uses common design patterns like selecting something or moving a slider, instead of trying to get a prompt just right. So I think that AI tools that rely too heavily on a user figuring out a prompt won't win compared to AI tools that just make it easier to fine-tune things in ways that feel familiar. So do I think that finessing prompts, a.k.a. prompt engineering, will go away? No, not at all. But do I think that we'll use more obvious and intuitive fine-tuning design patterns that might make finessing prompts less common? Yes, absolutely. But that's just my opinion. And let me know yours in the comments down below. Oh, and if your startup is looking to hire some talented designers, go check out Numi.tech. I'm currently part of their team and they're doing amazing work. And if you want to hear more of my hot takes on AI and designing software, go ahead and follow me on Twitter. I tend to follow anyone who calls me Batman. So just saying. And apparently YouTube's AI algorithm thinks that you should watch this video next for some reason. God bless. Batman out.
