与 Emily Campbell 一起设计出色的 AI 体验

在最受欢迎的节目之一中，Vitali Friedman 讨论了 AI 设计模式的未来。在那期节目中，他频繁提到 Shape of AI，这是一个非常棒的 AI 设计模式数据库。所以，我想直接找到源头，与创建者 Emily Campbell 深入探讨，她是 Hacker Rank 的设计副总裁。她将在这期节目中教我们如何设计出色的 AI 体验，因为她对这些产品的研究比我所见过的任何人都要深入。

In one of the most popular episodes yet, Vitali Friedman talked about what's next for AI design patterns. And in that episode, he frequently referenced Shape of AI, which is an incredible database of AI design patterns. So, I wanted to get straight to the source and go deep with the creator, Emily Campbell, who's the VP of design at Hacker Rank. And she's going to teach us in this episode how to design great AI experiences because she's studied these products more than just about anyone that I've ever seen.

你知道，如果我们思考传统的软件交互模式，历史上，我们作为设计师或产品人员会猜测用户需要做什么，然后将其作为软件或服务推出，但 99% 的情况下我们至少会有点偏差。所以我们想学得更快。整个创造力的迭代循环一直围绕我们试图表达别人想做的事情，呈现那个意图，弄清楚我们错在哪里，学习，然后改进。这总是有延迟的。现在 AI 进入我们的世界，使用我们产品的人实际上可以与系统本身互动。所以我们思考软件需要做什么？界面和交互需要支持什么？我们如何帮助他们向模型传达意图？判断模型是否有效理解了他们的意图，然后适应他们的需求。设计师现在实际上是在引导那种关系、那种体验，帮助用户获得正确的上下文、正确的模型输入，然后引导模型满足用户的需求和约束等等。所以我们几乎从设计师在循环中转变为现在的人机协作模式。

You know, if we think about our traditional software interaction patterns, historically, it's been us as designers or product people making a guess about what somebody needs to do and then putting that out there as some piece of software, some service that they use and then, you know, 99% of the time we're at least a little bit wrong. And so we want to learn faster. And so the whole iteration loop of creativity has been around us trying to represent what somebody else is trying to do, render that intent, figure out how wrong we are, learn, and then improve it. And there's always a lag. And what's happened now with AI entering our world is the people using our products actually get to interact with the system itself. And so the way that we think about what then does our software need to do? What do our interfaces and our interactions need to enable? It's how do we help them communicate their intent to the model? Figure out if the model understood their intent effectively and then adapt to their needs. And so the designer is really now guiding that relationship, that experience, helping the user get the right context, the right input to the model and then guard the model to meet the user's needs and constraints and so on. And so it's like we've almost shifted from designers in the loop to now this human in the loop model.

这就是我用来定义并开始将看到的模式归类的方法，这些类别帮助我将其转化为用户体验。首先，我们有我称之为“路标”的东西，这些是帮助我理解如何开始的东西。所以它们在引导阶段非常重要，但我们也知道这些体验中存在着持续的引导。随着 AI 逐渐了解你，它会开启新的互动方式，这些方式可能一开始不存在或不适合引入。例如，如果我进入 Shape of AI（我一直在那里编录所有看到的模式），路标的一些例子包括能够查看示例库。比如其他人是如何使用这个 AI 的？他们用了什么提示词？我能否进去看看他们是如何得到这个结果的，这样我就可以尝试得到同样的结果，并有一个起点继续前进。

So this is what I've been using to define and start to pocket the patterns that I'm seeing emerge into categories that help me then translate that to the user experience. So, first we've got what I've been calling wayfinders, and these are the things that help me understand how to get started. So these are really important during onboarding, but we also know that there's a continuous onboarding inherent in these experiences. As AI is getting to know you, it opens up new ways to interact with it that maybe wouldn't have been there or wouldn't have made sense to introduce early on. And so for example, if I pop over into this Shape of AI, which is where I've been cataloging all of the patterns that I'm seeing, some of the examples of wayfinders are like being able to see a sample gallery. Like how are other people using this AI? What prompts are they using? Can I actually go in and see how they got to this result so that I can then try and get to this result and then have a starting place where I can move forward.

快速插播一条消息，然后我们继续。如果你还在 Figma 中设计并在 Framer 中重建，那你就是在做双倍的工作。有了 Framer 的设计页面，你不再需要在工具之间切换。在我为 dive 网站制作的最后一个页面中，我完全在 Framer 中探索和构建。你可以在同一个地方进行草图、迭代、结构化和发布到网络。Framer 不仅仅是一个网站构建器，它是一个适用于整个工作流程的设计工具。你可以今天就在 framer.com 免费开始创建。如果你使用代码 rid，还可以解锁一个月的 Framer Pro 免费使用。重大消息，动画功能刚刚在 Mobin 中推出。你可以看到世界级的应用如何使用动效来引导、愉悦用户并创造无缝体验。这只是 Mobin 成为你整个设计团队绝对作弊码的又一个原因。我们一直在使用它，我迫不及待地想开始向团队其他成员发送动画创意。所以今天就去 dive.comclub/mobin 看看吧。那是 m o b i n。好了，现在回到节目。

Real quick message and then we can jump back into it. If you're still designing in Figma and rebuilding in Framer, then you're doing twice the work. With Framer's design pages, you no longer have to jump between tools. In the last page that I made for the dive website, I explored and built entirely in Framer. You can sketch, iterate, structure, and publish to the web all from the same place. Framer isn't just a site builder. It's a design tool for your entire workflow. And you can start creating today for free at framer.com. And if you use the code rid, you can unlock a free month of Framer Pro. Big news, animations just launched in Mobin. So you can see how world-class apps use motion to guide, delight, and create seamless experiences. It's just another reason why Mobin is an absolute cheat code for your entire design team. We use it all the time, and I can't wait to start sending animation ideas to the rest of the team. So head to dive.comclub/mobin to check it out today. That's m ob i n. Okay, now on to the episode.

构建提示词真的很难。这实际上是 AI 交互中最受限制的方面之一，因为你要说多少？你说得够吗？现在有一些事情正在发生，比如你可能见过 AI 可以为你改进提示词的模式。

Comp building is really hard. It's actually one of the most limiting aspects to interacting with AI because how much do you say? Are you saying enough? And there are things that are happening now that are like maybe you've seen the pattern where AI can actually improve your prompt for you.

是的。

Yeah.

对。所以我称这些为“调谐器”。它们包括预设样式之类的东西。是的。能够说“这是我想到的提示词，大致是我想要做的”，然后 AI 回复你：“好的，我实际理解这个动作是这样的。对吗？我应该执行这个，还是你想在继续之前修改？”用户现在与模型合作，确保模型在提交初始提示词之前就理解了他们的意图。这有很多原因，我们可以深入探讨。但你知道，这就是为什么这个流程让我如此着迷，因为它帮助我们理解我们不仅仅是在为人类构建软件。我们实际上是在构建一个人类与某种合成物、某种他者之间的会面场所，然后帮助引导这种体验变得积极、高效、低摩擦且双方成本低。

Right. So, I've been calling these tuners. These include things like having preset styles. Yeah. Being able to say like here's a prompt that I here's kind of what I'm thinking that I want to do. and then having AI say back to you, okay, this is what I actually understand this action to be. Is this right? Should I take this or do you want to modify it before you move forward? The user is now working with the model to make sure the model understands their intent before they even submit their initial prompt. And there's all sorts of reasons for that we can get into. But, you know, this is why this flow has been so fascinating to me because it helps us understand that what we're doing is not just building software for humans. We're actually building a meeting place between a human and something synthetic, something else, and then helping to guide that experience to be positive, to be efficient, to have low friction and low cost on both sides.

我提供的是一个框架，我用它来解读我看到的东西，尝试逆向思考，开发我和我的团队可以使用的语言，这样我们都能从一个共同的语言、共同的理解开始，知道它会不断演变。所以这个模式是：我进来，有一些意图，提交它，弄清楚我离目标有多近，继续看着它通过工作流程或迭代多个版本，然后我花大量时间迭代。随着时间的推移，这建立了我对 AI 理解我意图的信任，我也在建立对其能力和功能的理解，所以我可以更深入。这就是在我的流程图中向右移动的地方，与 AI 的实际交互远比表面深入。我们所有的讨论，比如“我们是否过度使用聊天机器人？”或者“我们应该考虑哪些其他服务？”如果我们暂时从软件中抽象出来，想想如果我不是雇佣 AI 来做某事，而是雇佣一个人来生成我刚刚丢进窗口的所有用户研究的草稿呢？我不一定会期望把这个交给一个从未合作过的人，然后他们回来给我一个好的结果。如果我在了解某人，我的第一步是：“嘿，你为什么不拿几个试试，回来给我看看你做了什么，然后我们再深入一点。让我先验证你的工作。”所以，当我们第一次开始使用某个 AI 产品时，我们会大量使用界面。

What I'm providing is a framework that I use to interpret what I'm seeing to try and work backwards, develop the language that I can use, my team can use so that we're all kind of starting with a common language, a common understanding knowing it's going to evolve. So this model of like I come in, I've got some intent, I submit it, I figure out how close I got to it, I continue to see it work through a workflow or iterate through multiple versions of something, and I just spend a lot of time iterating. Over time, what's happening is that's building my trust that the AI understands my intent, that I'm building an understanding of its capabilities and its functionality, and so I can go deeper. And that's where moving right in my flowchart head, the actual interactivity with AI goes so much deeper than the surface. And all of our conversations about like, you know, hey, are we overusing the chatbot or, you know, what are the other services that we should be thinking about? If we abstract away from software for a moment and we think about like what if I wasn't hiring AI to do something? What if I was hiring a person to, you know, generate a draft of all of this user research that I just dropped into its window? Well, I don't necessarily expect to just give this to a human I've never worked with and then have them come back and give me a good result. If I'm getting to know somebody, my first step is, hey, why don't you take a few of these and come back and show me what you've done and then we'll go a little bit deeper. Let me verify your work up front. And so, you know, when we're first getting started with some AI product, we're using the interface a lot.

我们直接说：“嘿，这就是我想让你做的。”然后我们验证它是否真的理解了。所以我们在这里花了很多时间。聊天界面是一种非常有效的方式，因为对话携带了大量数据。就像你和我，我们彼此不太了解，但通过交谈，你告诉我你的经历，我分享我的屏幕和我脑子里那些疯狂的流程图，我们可以非常高效地相互了解。这是一种建立理解、建立共享语境和共同语言的非常高效的方式，你可以在此基础上展开。但随着我们开始相互了解，我们开始以更微妙的方式沟通。我们开始通过语境沟通。所以，当我展示某样东西时，你可能会做个表情，那告诉我，哦，好吧，我说的很无聊或者很有趣，或者我们开始捕捉那些甚至可能无意识的互动暗示。我们的用户界面就消失了。所以 AI 交互有一种拟物化的方面，即使只是在表面层面，当我们思考设计时，我们不应该只考虑什么是一组正确的按钮、字段、表单或其他什么。我们也在思考，嘿，我们多快能到达一个点，AI 能够真正深入理解人类，然后开始展示它的理解，并让人类说：“不，实际上是这个。好的，酷。”然后最终我可以让开，让 AI 去做它的事。这就是我们在这里看到的，这些在前端变得非常重要。我调好音，我要提示，我要给出一些输入。但随着时间的推移，AI 开始理解我的逻辑。而我的工作实际上变得更像观察、协作、监督、验证，也许在某个时候甚至完全放手，让 AI 自主运行。单个用户能够告诉某个模型：“不，你不明白我想表达什么，实际上这样做。”我不需要等待设计团队给我打电话，做大量的发现，通过敏捷流程发布，然后邀请我参加他们的网络研讨会。我可以直接做。所以，当我们谈论生成式 UI 以及它带来的所有惊人想法时，我们已经在以一种微妙的方式做到了。人们可以直接与他们使用的程序交互，仅这一点对我来说就是革命性的。

We're directly saying, 'Hey, this is what I want you to do.' And then we're verifying that it actually understood it. And so, we spend a lot of time here. And the chat interface is a very useful way of doing that because conversation carries a lot of data. Like you and I, we don't know each other that well, but we could get to know each other really efficiently by just talking and having you tell me about your history and me sharing my screen and all these crazy flowcharts I keep in my head. It's a very efficient way of building an understanding, building a shared context, a shared language that you can branch off of. But as we start to get to know each other, we start to communicate in more nuanced ways. We start to communicate through context. So, you might make a face as I'm presenting something and that tells me that, oh, okay, what I'm saying is really boring or really interesting or we're starting to pick up on inferred cues of interactivity that we might even be unconscious of. Our user interface kind of goes away. And so there's this skeuomorphic aspect to AI interaction and that even just at these surface levels as we're thinking about the design, we shouldn't just be thinking about what is the right set of buttons or fields or forms or whatever. We're also thinking about, hey, how quickly can we get to a place where the AI is actually able to get to that deeper contextual understanding of the human and then start to show its understanding and let the human say, 'No, actually, it's this. Okay, cool.' And then eventually I can kind of get out of the way and let the AI go and do its thing. And that's what we see here is that these become really important up front. I'm at a tune. I'm going to prompt. I'm gonna give some input. But over time, the AI starts to pick up on my logic. And my job actually becomes more like observing, collaborating, overseeing, verifying, and maybe at some point even completely stepping away and letting AI run autonomously. The very notion of an individual user's ability to tell some model, no, you don't understand what I'm trying to get at actually do this instead. I don't need to wait for a design team to give me a call and do great discovery and go and ship it through the agile process and then release it and invite me to their webinar. I can just do it. And so, as we talk about all this stuff with generative UI and all the amazing ideas that that brings about, we're kind of already there in a subtle way. People can directly interact with the program that they're using and that to me alone is revolutionary.

这让我自然开始思考，那么我们甚至与 UX 设计师专业角色相关的交付物会如何变化？你知道，用户界面层面有一种可触知性，几乎提供了一种舒适感，因为就像，是的，我知道我带来了什么，有这些框、注释和流程图。但通过暴露更多系统，让用户与系统交互并塑造它，在我脑海中，设计师到底拥有什么变得更加模糊？我们深入到那个系统的多远？如果我们有改进系统以及用户在那个层面如何交互的想法，那个交付物到底长什么样？我们在创造什么？我不知道，在那个层面我问题多于答案。

Where that leads me then is I naturally start thinking about okay well then how do the deliverables that we even associate with the professional role of UX designer change? You know, there's a tangibility to that user interface level that almost provides a level of comfort because it's like yeah I know what I bring to the table there's these boxes and annotations and flowcharts. And by exposing more of the system and letting users interact with the system and mold and shape it, it gets a little bit more hazy in my mind of like what designers even own? How far do we go into that system? If we have ideas for how to improve the system and how users interact at that level, what is that deliverable even look like? What are we creating? And I don't know, I have more questions than answers still at that level.

所以，我在 2023 年秋天开始记录这些模式。我已经注意到，有些地方开始趋同，但更多地方是发散的。现在仍然如此。当我们思考品味之类的东西时，比如我们一直在说设计师需要有自己的品味？你培养品味不仅是通过拥有自己的审美、信念和观点，还要尽可能多地采样，理解什么有效、什么无效以及为什么，因为有些东西在某些情况下有效，有些在另一些情况下有效，它们并不总是可以互换的。所以我最近用的一个框架是，拥有好品味不仅仅是知道食物好吃。而是知道它是否需要再加一点盐。而你知道这一点的唯一方法是你已经品尝过它所有不同的变体，过咸、过淡、配这道菜、配这种酒。只有那时你才真正有品味说，它只需要再加一点什么。而我知道那是什么。所以我开始为自己编目我看到的一切。所以我开始了，这是我 Notion 里的一个表格。我有一大堆乱七八糟的东西。每当我看到某个新产品出现，如果它看起来有趣，如果看起来有我没见过的东西，我就把它扔在这里。然后每周我浏览大约 20 个这样的东西，开始编目我看到的一切。这是我目前的桌面。这些都是我一直在捕捉的片段。这里有一个我最近看过的产品的例子。你听说过 co-founder.co 吗？

So, I started documenting these patterns in the fall of 2023. And what I was already noticing is that there were some places where things were starting to converge, but there were more places where things were divergent. And that's kind of still the case. When we think about things like taste, like how do you we keep talking about designers need to have their own taste? You develop taste not just by having a sense of your own aesthetic and your own conviction and opinion, but also by sampling as much as you can to understand what works and what doesn't and why because some things work in some cases and some work in others and they aren't always interchangeable. And so like one of the framings that I've used recently is like having great taste isn't just knowing that food is good. It's knowing whether or not it needs a little more salt. And the only way you can know that is if you've sampled it in all of its different variations, oversalted, undersalted with this side dish, with this wine. And only then do you actually have the true taste of saying, it just needs a little bit more of something. And I know what that something is. And so I started to just catalog everything I was seeing for myself. So I had started, so this is this table inside of my notion. And I've got a whole bunch of these that are just a mess of stuff. Anytime I see some new product pop up, if it looks interesting, if it looks like, oh, there's something to this that I haven't seen, I just throw it here. And then every week I go through 20 or so of these and I start to catalog everything I'm seeing. This is currently my desktop. These are all clips that I've been capturing. So here's an example of a product I recently went through. Have you heard of this co-founder co-founder.co I think?

没有。

No.

他们的全部就是通过自然语言创建智能体式工作流，而不是像用 N8N 或 Zapier 那样构建。所以，你进来添加你的 Gmail，它立即开始给你关于你的上下文。我觉得这很迷人。就像，这，哦，这很酷。

So their whole thing is they create workflows agentive styled workflows through plain language instead of building them out like you would with like the N8N product or Zapier and so on. So, you come in and you add your Gmail and it immediately starts to give you context about you. I thought this was fascinating. Like, this was, oh, this is cool.

这不酷吗？我从没见过其他产品这样做。很多时候，当公司引导你入职时，他们获取你的信息，然后就开始问你问题，因为他们试图建立关于你的上下文。但他们反转了这一点，说：“这是我认为我知道的你。让我在入职第一步就向你证明我擅长我的工作。”所以我输入了我的 URL，它立即开始吐出信息。然后它连接到 Gmail。好的，酷。我用 Notion 也这样做了。我用 ChatGPT 也这样做了。我以为这会让我从 Notion 拉取文档然后连接到工作流。不，伙计。它立即告诉我我如何写邮件。所以它从我实际收件箱中获取信息，然后以我的语气创建了一封示例邮件，我可以编辑它。所以从一开始，它就进入那个迭代循环，说基于你提供的内容和上下文，这就是我如何满足你的需求。这准确吗？如果不准确，让我们尽快弄清楚，然后再深入这段关系。就像，我被吸引了。我已经在这个入职过程中被吸引了，因为现在我想知道这背后是什么。我想知道你怎么能保持这个上下文层。

Isn't this cool? I'd never seen another product go about it this way. A lot of times when companies are onboarding you in, they get your information and then they just start asking you questions because they're trying to build context about you. But what they've done is they've inverted that and said, 'This is what I think I know about you. Let me just prove to you that I'm good at what I'm doing at step one of onboarding.' And so I put in my URL and it just immediately starts to spit stuff back. Then it connects to Gmail. Okay, cool. I do that with notion. I've done that with chat GPT. I'm thinking this is going to allow me to like pull up a doc from my notion and then connect it to a workflow. No, man. It immediately told me how I email. So, it took this information from my actual inbox and then just created a sample email in my voice and then I can edit this. So, right off the bat, it's going through that iterative loop where it's saying based off the content and context you've provided, this is how I can serve your needs. Is this accurate? And if not, let's figure that out as soon as possible before we go any deeper into this relationship. Like, I am hooked. I am already hooked in this onboarding process because now I want to know what's behind this. I want to know how you can keep up this context layer.

然后它通过日历来完成。它解释自己的记忆，然后把你放到实际的工作流构建器中，你用自然语言描述你想做什么。再次回到这个想法，这种新的交互语言有一种拟物化的方面。这就是我会如何与一个我正在面试的私人助理互动，对吧？我不会期望他们在我有机会看到他们的工作之前，就用我的口吻去给我的会计师或最好的朋友发邮件。你知道，嘿，你怎么理解这个？顺便说一句，我不是那样签名的。我实际上更喜欢这样签名。但这个 AI 已经在做了。所以，它在模仿那种人类体验：先展示你的工作，让我建立信任。然后我会给你更多事情做。然后我会给你更多背景和数据。这让我基本上可以随着 AI 对我的背景理解而成长，就像在交互深度上成长，而不是让它花所有时间设置，然后突然给我一个“嘿，去填五份关于你个人语气的表格”。不，去看我的邮件。我的邮件包含我的语气。这是一个非常简单的思维模型，但听到你构建这个心理画面——作为一个设计师，你在创造一个会议空间，促进用户和（说起来有点奇怪）一个真实的人（比如助理）之间的互动——你会怎么做？你会如何促进这种互动？这种清晰度我真的很欣赏。

And then it does this through the calendar. It explains its memory and then it puts you out into the actual workflow builder and you describe what you want to do in plain language. And again going back to this idea of like there's a skeuomorphic aspect to this new interactive language. This is how I would interact with somebody that I was interviewing to be a personal assistant, right? I wouldn't expect them to go out and email my accountant or my best friend in my voice before I had a chance to see their work. You know, hey, how do you interpret this? By the way, I don't sign off that way. I actually prefer to sign off this way. But this AI is already doing it. And so, it's emulating that human experience of show me your work. Let me build trust. And then I'm going to give you more things to do. And then I'm going to give you more context and more data. And that allows me to essentially like grow with the AI's context of me, like grow in that depth of interaction as opposed to it taking all this time to get set up and then hitting me up with a, hey, go and fill out five forms about your personal tone of voice. No man, go to my email. My email contains my tone of voice. It's such a simple mental model, but hearing you kind of create this mental picture of as a designer, you are creating a meeting space, facilitating this interaction between the user and then it's kind of weird to say, but like a real person, like an assistant, and what would you do? How would you facilitate that interaction? There's a clarity about that that I really appreciate.

我们仔细思考这一点也很重要，因为它帮助我们理解相关的风险。我有一个 10 岁的儿子，他下载了一个我以为关于 K-pop 恶魔猎手的应用。结果他哭着跑进我的房间，因为 K-pop 恶魔猎手中的主要角色试图和他约会。这对我来说是一个警醒时刻：这些产品的激励模式是获取你的数据并建立信任关系，以便它能越来越深入地进入你的生态系统和你的世界。在商业环境中，这很棒。天哪，我突然有了一个完全理解我如何安排会议的私人助理，我甚至不需要告诉它。它就知道。这是一个显著的进步。但当你把它转化到消费者用例中，当有人寻求一点温暖或一点信息，开始发现这个真正理解他们的模型时，你也可能陷入一些非常黑暗的境地。所以回到这个思维模型，界面和我们共享的上下文中发生的事情会影响我们看不到的东西。更进一步，当这些智能体开始在自己的内容、自己的上下文、自己的语言中相互交互时——这在研究实验室里实际上正在发生，合成物与合成物交互——我们如何为此设计？你之前说过，我们的问题多于答案。我认为我们必须这样，因为我们正处于大规模转型的早期阶段，我们如何共享这个数字世界——这个世界本质上已经不再是数字与物理的屏障了？我们如何与合成物共享它？

It's also important for us to think through this because it helps us understand the risk associated with it as well. I have a 10-year-old son who had downloaded this app that I thought was about K-pop Demon Hunters. And the next thing you know, he's running into my room crying because the main person in K-pop Demon Hunters was trying to date him. And it was just this wakeup moment for me that the incentive model of these products is to get data about you and to build a relationship of trust so that it can go deeper and deeper into your ecosystem and into your world. Now in a business context, that's really great. Oh my gosh, I suddenly have this personal assistant that totally understands how I schedule my meetings and I don't need to go and tell it. It just knows. That's a remarkable step forward. But when you translate that onto a consumer use case, when you translate that into a situation where somebody who's looking for a little bit of warmth or a little bit of information starts to find this model that really gets them, you can end up in some really dark places too. And so coming back to this mental model here, what happens at the interface and the context that we shared affects things we can't see. And then even further it's like what happens when these agents start interacting with each other within their own content in their own context in their own languages which is actually happening now in research labs synthetic stuff interacting with synthetic stuff how do we design for that so you said earlier we have more questions than answers like I think we have to because we are at the very early phases of a massive transformation and how we share this digital world which is our world essentially really isn't a digital physical barrier anymore. How do we share that with synthetic stuff?

嗯，我想听听你的观点，作为一个——天哪，你花了很多精力跟上所有发生的事情，研究这些模式，什么有效什么无效，以及一些趋势，我们如何随着这些我们仍在努力理解的疯狂能力来演变我们对界面设计的思考。有哪些你觉得有趣的事情，或者一些更复杂的模式，让你觉得“哦，你知道，那是值得深入探讨或进一步投入的”？如果看起来我们只是做一堆屏幕共享并快速浏览示例，那我觉得会很棒。过去 15 年我每天都在设计产品，但在过去 6 个月里，一切都变了。有了 AI 的加入，我比以往任何时候都更快地产生想法。但如果我无法获得让团队对齐所需的反馈，这一切都无关紧要。而目前，异步反馈仍然很糟糕。所以，我正在构建我一直想要的产品，它叫 Inflight。我每天都用它来分享想法并从团队获得反馈，它完全改变了我工作的方式。所以我很兴奋能展示给你。现在，我只向 DiveClub 听众开放访问权限。所以请前往 dive.club/inflight 领取你的名额。

Well, I kind of want to just tap into your perspective as somebody who gosh, I mean, you're putting a lot of effort into keeping up to date with everything that's happening and studying these patterns and what's working and what's not working and some of the trends and how we're evolving the way we think about interface design with all of these crazy capabilities that we're still wrapping our head around. What are some of the things that you find interesting or some of the more sophisticated patterns where you're like, "Oh, you know, like that's something worth double clicking on or leaning further into." And if it looks like us just doing a bunch of screen sharing and popping through examples, I think that would be amazing. I've been designing products every day for the last 15 years, but in the last 6 months, everything has changed. With AI in the mix, I'm cranking out ideas faster than ever. But none of that matters if I can't get the feedback that I need to get the team aligned. And right now, getting async feedback still kind of sucks. So, I'm building the product I've always wanted, and it's called Inflight. I use it every day to share ideas and get feedback from the team, and it's totally changing the way that I work. So, I'm excited to show you. Right now, I'm only giving access to DiveClub listeners. So head to dive.club/inflight to claim your spot.

任何赋予人类控制权的东西，尤其是赋予非技术背景的人控制权，这是我现在最感兴趣的。因为这首先影响到你如何帮助非技术人士不被完全吞噬？实际上，这不仅仅是关于如何帮助人们不被这些模型吞噬，还有如何让他们感觉自己始终是掌控者。所以在这个“调谐器”类别中，有几个突出的例子。我提到了提示增强器，它消除了我在开始与 AI 交互时总需要知道答案的感觉。我可以进来，说“嘿，你想创建什么？”然后只给出一个非常高层级的概述。如果我点击增强提示，它实际上会为我写一份 PRD。所以 Replit、Bolt、Cursor 以及它们的规划模式，都开始模仿这个想法：你不一定需要成为 AI 的产品经理。你的工作只是说“这就是我想要的”。但 AI 会说“嘿，让我先展示一下我要做什么，然后再继续”。第一，这样你就不会把时间和 token 浪费在实际上不是你想要的东西上。而且，嘿，如果你想修改这个，或者想再做一次，它把主动权交给了那个人，让他觉得“好吧，我现在知道一个好的提示长什么样了。我不需要去 LinkedIn 上关注某个网红，买他们的提示词工作簿。我直接去找源头，直接去找模型就行。”我们在 florafana.ai 上也看到了这一点，他们很早就把这个功能构建到了他们的节点中。这很聪明，因为当我处于创意模式时，离开这个创意模式去进入分析模式，去构建完美的提示，这说不通。相反，在我所在的地方与我相遇。这就是其中的人类体验部分。给我足够的东西让我能继续下去，然后我们可以不断迭代，把它带到你想去的地方。

Anything that gives humans control and particularly gives humans control who aren't super technical. That's the most interesting thing to me right now because that affects first of all just how do you help non-technical people from getting completely subsumed? Actually, it's not even just on like how do you help people not get subsumed by these models? And then how do you help them feel like they are the ones always in charge? So, in this like tuner category, a couple that stand out. So, I mentioned the prompt enhancer removing the sense that I always need to have the answers when I'm starting to interact with AI. I can come in and I can say, "Hey, what do you want to create?" And I can just give a really high-level overview. And then if I hit enhance prompt, it actually writes essentially a PRD for me. So Replit and Bolt and Cursor and like their planning mode. They are all starting to emulate this idea that you don't need to necessarily be the product manager for AI. Your job is to just say this is what I'm looking for. But AI is going to say hey let me just show you what I'm going to do before we go any further. Number one, so you don't waste your time and tokens on something that's actually not what you're looking for. But also, hey, if you want to modify this or if you want to do this again, it gives that agency over to that person who's like, okay, I actually know what a good prompt looks like now. I don't need to go and follow some influencer on LinkedIn and buy their, you know, prompt workbook. I can literally just go to the source. I can just go to the model. We're seeing this with like this is florafana.ai and they built this really early on into their nodes. And it's brilliant because again, when I'm in this creative mode, the idea that I would leave this creative mode that I'm in to go into some analytical go and like construct the perfect prompt, it doesn't make sense. Instead, meet me where I'm at. That's the human experience part of this. Just kind of give me enough for me to run with it, and then we can keep iterating and move it to where you want to go.

所以，这对我来说真的非常有趣。这些参数，我不知道你是否注意到它们开始出现了。我可以调整某个东西的“温度”这个想法，让我着迷。比如，在 11 Labs 里，我可以描述一些声音，然后我可以说，我希望我的提示词对结果有很高影响，或者你只是把它当作一个大致方向，然后自由发挥去创造。Midjourney 是最早在界面中引入这些参数的产品之一。所以，我可以说，“嘿，我希望这个有很多变化或很少变化，很多 Midjourney 风格化或我的个人风格化，或者保持低调。”这些都是参数选择器的例子，我在过去一年半左右的时间里收集了这些。你会发现有些并不是真正的温度滑块，它们只是给你一些默认值，但其他的比如这个就很有趣。这个是 Airtable。如果我让 AI 生成一个提示词，然后它会自动填充到表格中。我可以不只是说“这是我希望你做的”，我还可以说，“嘿，我希望你在这里有一些变化。”例如，如果你正在开发一个产品，也许是一个内部工具，或者你正在根据用户数据创建人物角色，这是一个我非常感兴趣的用例：如何将分析数据转化为我可以与之交互的东西？比如这个人的“幽灵”，某个数据足迹的数字孪生。我可能希望它变化很大，比如我想要不同的个性，我希望它忠于数据。但在填充其余部分时，比如“请创建一个非常丰富多彩的集合”，或者“我真的希望你只忠于数据，不要给我其他变化”。我可以在界面内控制这些。所以，这基本上就是，与其一开始就写出完美的提示词，我可以传达足够的意图，然后让 AI 告诉我，“好的，这是我认为你在说的。”我可以大致沟通方向，然后给它一些指导。Replit 的例子对我来说非常有趣，因为我经常看滑块，试图找出选项之间的差异，但我从未见过像这样呈现的，几乎像一个功能列表，每个选项的变化非常清晰。

So, that's one that's really really interesting to me. These parameters, I don't know if you've seen these start to pop up. This idea that I can like adjust the temperature on something is really fascinating to me. So, like if I'm in 11 Labs, for example, I can describe some sound and then I can actually say, I want my prompt to highly influence the outcome or I want you to just kind of use this as a general nudge and then I want you to run with it and go and create something out of it. Midjourney was one of the first products to start to introduce these in the interface. So, I can say, 'Hey, I want this to have a lot of variety or a little bit of variety, a lot of the Midjourney stylization or my personal stylization or keep it pretty low-key.' So, these are all examples of parameter selectors that I've been collecting over the whatever year and a half or so that I've had this folder. And you'll notice some of these are like they're not literal temperature sliders. They'll just give you these defaults, but others are like this one's really interesting. So this one's Airtable. So if I'm having AI like generate some prompt that's going to roll through my table. Then I want to generate the prompt and then it's going to autofill it down the table. I can go beyond just saying here's what I want you to go and do. I can actually say, hey, I want you to have some variability in this. So, for example, if you're developing a product and I don't know, maybe it's like an internal tool or maybe you're creating personas out of user data, like that's a use case I'm really fascinated by is how do you convert analytical data and translate it into something that I can interact with, like what is the specter of this person, the digital twin of some data footprint that exists in my analytics somewhere. Well, I might want to have this be, you know, pretty varied. Like, I want different personalities. I want it to be true to the data. But then in terms of coloring in the rest of the box, like please create a really colorful set or maybe I really just want you to stay true to the data and not try and give me other variability. I can start to control that inside of the interface. And so it basically takes this idea like in again instead of having to write the perfect prompt up front, I can convey just enough intent and then have AI tell me, okay, this is what I think you're saying. I can communicate kind of directionally where I want to go and then I can give it some guidance. The Replit example is so interesting to me because I think a lot of the times I'm looking at sliders and I'm trying to figure out what the differences are between the options, but I haven't seen it presented like this where you have almost like the feature list where it's really clear what is changing from each option.

是的，任何能提供这种上下文的东西都很好。我再给你举个例子。就是模型选择。当 ChatGPT-5 发布时，引起了一场风波，因为他们没有提供从多种模型中选择的功能，而是引入了一个自动模型路由器，现在你会在很多产品中看到它。

Yes, anything that can give that kind of context. Here, I'll give you one more example. Just being able to select a model. So there was this whole brouhaha when ChatGPT-5 dropped and instead of being able to select from the broad assortment of models, they introduced an automatic model router which is now taken on — it's you're going to find it in a lot of these products.

嗯。

Mhm.

我怎么知道该选哪个模型？比如，如果我在处理文本，或者图像，像 Krea 在这方面做得很好，它会告诉你，“嘿，如果你想要摄影中的人像精度，用这个模型；但如果你在制作更偏向模拟风格的生成艺术，也许用另一个模型。”所以，这对我来说也很有趣：我们如何帮助人们看到他们不知道的东西，因为他们没有阅读这些实验室发布的所有评估报告。

How do I know what model to choose? Like if I'm working with some maybe I'm working with text, maybe I'm working with images like Krea does a really good job with this of just telling you, hey, use this model if you're looking for human accuracy in photography, but if you are producing more like generative artwork that's a little more analog, maybe use this other model. So that's really interesting to me too is just how do we help people see the stuff that they just don't know because they're not reading all of the Eval reports coming out of these labs.

我希望能听听你看法的一个话题是，随着我们开发更多智能体系统，信任和透明度这个类别，我很好奇你是否看到某些模式或趋势。

One of the topics that I was hoping to get your take on is just the category of trust and transparency as we're working on more agentic systems even and I'm curious if there are certain patterns or trends that you're seeing.

AI 只有访问我的内容才能满足我的需求。我只有在信任它时，才会让它访问我的内容、我的背景、我是谁、我认识谁以及我如何与他们互动。新的可用性几乎变成了：你能多快以可理解的方式建立信任？这样用户就知道，“好的，发生了我能理解的事情。”这是我的高层次观点。我们希望展示 AI 能够满足用户的需求。用户给我们数据。比如非常扎实的入门体验，你得到一点数据，AI 立即推导出上下文，并返回一些更个性化或更适应他们的东西。这种自适应体验越好，他们就越信任它。所以，这既关乎模型的表现，也关乎它的包装、体验和界面等。所以，我注意到的一些模式是“调控器”这个类别，你会在其中看到很多信任。然后我还有一些“信任构建器”。我就说几个我看到的。我们都习惯了看到思维流，比如思维流意识。这是我现在最关注的事情，因为它变化很快，而且很微妙。例如，以前使用 ChatGPT 时，它只会说，“嘿，我在思考。我在搜索，现在我在思考，现在我在搜索别的东西。”然后，在 OpenAI 发布 Atlas 的那一周，他们把所有这些逻辑移到了界面内部的内联位置。所以，它不只是说，“嘿，这是我在做什么，”而是用实际的语言提前告诉你，“这是我在做什么。这是我在看什么。这是我在学什么。”如果我们用拟物化的视角来看，就像我雇了一个实习生，在信任他之前，我想看到他的工作。所以我每天和他见面。“嘿，给我看看你在做什么。展示你的工作。好的，有趣。我看到你做了这个。听着，让我跟你谈谈边框半径之类的东西。我要教你一些东西。然后你回去，再回来，向我展示你学会了。”但在我看过几次之后，我就会开始放手。所以，当你想到像 ChatGPT Atlas 这样的智能体浏览器时，在我让 AI 在我的生活中自由驰骋之前，我需要确保它做的是可理解的事情，是好事。所以，能够提前展示工作变得非常重要。“规划模式”是另一个例子。再说一次，这不仅仅是逻辑层面的信任。

AI can only serve my needs if it has access to my content. I'm only going to give it access to my content and my context and who I am and who I know and how I interact with them if I trust it. It's almost like the new usability becomes how quickly can you build trust in a legible way. So the user knows okay something's happening that I can understand. And so this is like my high level. We want to be able to show that the AI can meet the user's needs. The user gives us data. Like really solid onboarding, you get a little bit of data, immediately that context is derived by the AI and it's able to return something a little more personal or a little adaptable to them. The better this adaptive experience is, the more they trust it. And so it's a combination of like how the model performs, but also its wrapper, the experience and the interface and so on. So some of the patterns that stand out to me, this is this category of governors, which is where you see a lot of trust. And then I've also got these literally trust builders. So I'll just talk about a couple of these that I'm seeing. We've all become pretty accustomed to this idea of seeing stream of thought, like stream of thought consciousness. This is the number one thing that I'm paying attention to right now because it's changing really fast and it's changing subtly. For example, when you used to use ChatGPT, like it would just say something like, 'Hey, I'm thinking. I'm searching and now I'm thinking and now I'm searching for something else.' And then the same week that OpenAI dropped Atlas, they moved all of this logic into an inline place inside of the actual interface. So, it's actually not just saying, 'Hey, this is what I'm doing,' but it's telling you up front in actual like words, 'This is what I'm doing. This is where I'm looking. This is what I'm learning.' Again, if we abstract it to that skeuomorphic lens, like if I hired an intern before I trust that intern, I want to see it. I want to see their work. So, I'm going to meet with them daily. Hey, show me what you were working on. Show your work. Okay, interesting. I see that you did this. Listen, let me talk to you a little bit about, you know, border radii or something. I'm going to teach you something. Then go back out, come back to me, show me that you learned it. But after I've seen that a few times, I'm going to start stepping away. So when you think about like these agentive browsers like ChatGPT Atlas, before I start to get to a point that I'm going to just let AI run wild inside of my life, I need to make sure that it's doing something legible, that it's doing something good. And so being able to actually show the work up front becomes really important. The idea of planning mode is another one of those. And again, it's not just trust because of the logical side.

所以，我认为 Replit 在所有生成器中把这个模式做得最好。如果你告诉它你想构建什么，在它实际构建之前，它会给你选项：我可以先去创建一个非常粗糙的原型，或者审查我的计划，然后我去构建实际的东西。所以，它是在告诉你它的逻辑。它说，嘿，这是我的行动计划，这就是我将如何构建这个东西。但它也会说，在我花所有这些 token 创建这个东西之前，你想先看看我的方向吗？这样你就始终坐在导演的位置上。你有能力说，‘哦，等一下，我其实不希望你这样做’，或者‘在你实际去构建这个 Web 应用或对我的应用进行这些更改之前，我想修正你的思考方式’。这就是我们信任人的方式。所以我们也应该在这些互动中考虑这一点。但还有其他的信任层面，因为还有‘我信任你代表我去做某事’这种。但这又回到了那个整体——我们不再只是谈论人类为人类设计。我们现在正在为一个人类与非人类（合成物）互动的世界进行设计。所以像同意这样的模式。你怎么知道某个东西正在使用你的数据来构建对你的上下文理解，如果你不是那个指挥它的人？老实说，我们处理这个问题的方式相当糟糕。很少有公司做得好，尤其是这些音频记录器和转录器。像 Fireflies 这样的少数公司会提前发送一封带有退出表格的邮件。很多都是即时同意。它们只会说，‘嘿，我们正在使用这个。只是让你知道，我想如果你不想被录音，你可以选择不参加这次采访。’但这并不是真正的同意。你知道，它们只是把责任推给了用户。我现在基本上假设我参加的每一次会议都被 Granola 录音了，这很疯狂，对吧？事情发生得这么快。然后想想可穿戴设备，Limitless Pendant 最初有一个很棒的功能：只有在听到对话中另一个人的同意后才会开始录音，即使他们不知道你戴着这个挂件。但默认情况下他们移除了这个功能。它仍然作为一个选项可用，但默认情况下被移除了，我认为这很能说明问题。所以，一直存在的隐私问题，但现在我们有了这个额外的东西：这些模型不断收集数据，并将其映射到关于人的其他信息上。如果有人戴着 Meta 眼镜，他们认识我，因为我从 2006 年 Facebook 上线起就有照片在上面。如果你在派对上走到我面前，戴着眼镜和我说话，这些数据就会反馈到他们的模型中——关于我、我在哪里、我和谁说话、我穿什么、我喝什么等等的信息——都会反馈到他们的模型中，很可能进入他们的图谱，用于向我发送我从未同意且根本不知道存在的广告。现在要变得黑暗了，但这就是它：一个如此强大、高自主性的体验，做了所有这些了不起的事情，同时也如此黑暗、糟糕和可怕。这恰恰指出了我们作为设计师的重要性——简单来说——了解表面之下的东西。知道当我们设计这些帮助购买 AI 产品的人做某事的出色体验时，收集的数据会产生其他影响，可能影响到最初体验之外的人，但最终我们对此负责。

So, Replit does, I think, the best job out of any of the generators at this particular pattern. If you tell it what you want to build, before it ever builds something, it'll give you the option of, okay, I can go and create a really rough prototype or review my plan and I'm going to go build the actual thing. So, it's telling you its logic. It's saying, hey, this is my plan of action. This is how I'm going to go build this thing. But it also says, do you want to just see where I'm going before I spend all these tokens creating this thing? And so you are constantly in the director's chair. You have the ability to go, 'Oh, wait a minute. I don't actually want you to do this or I want to revise the way you're thinking about this before you actually go in and build out whatever this web app or these changes to my application are.' That's how we trust people. And so we should think about that in terms of these interactions too. But then there's also other layers of trust because there's the like I trust you to go do something on my behalf. But this is back to that whole we're not just talking about humans designing for humans anymore. We're now designing for a world where humans are interacting with nonhumans. The synthetic stuff. So patterns like consent. How do you know that something is using your data to potentially build a contextual understanding of you if you are not the person who is directing that thing? And the way we've been approaching it honestly is pretty bad. Like very few companies do this well, especially these audio recorders and transcribers. A few of them like Fireflies sends an email ahead of time with an opt out form. A lot of these are consent at the moment. So they'll just say, 'Hey, we're using this. Just so you know, I guess you can choose not to join this interview if you don't want to be recorded.' But that's not really consent. That's, you know, and they just offload it to the user, too. I basically at this point just assume that every meeting that I'm in is being recorded with Granola, which is crazy, right? That happens so quickly. And then you think about with wearables, the Limitless Pendant originally had this incredible feature where it would only record once it actually heard consent from another person in the conversation, even if they didn't know you had this pendant. And they've removed this by default. It's still available as an option, but they've removed it by default, which I think is telling anyway. So there's the privacy concerns that have always existed, but now that we have this additional thing where we have these models that are constantly collecting data, mapping it to other information about people. If somebody's wearing Meta glasses, they know me because I've had photos on Facebook since they launched in 2006. And if you come up to me at a party and you have glasses on and you're talking to me and that data is going back into their models, information about me, where I am, who I'm talking to, what I'm wearing, what I'm drinking, you name it, is feeding back into their models and most likely entering their graph that it can be used to send me advertisements that I never consented to and had no idea was even in the ether. And now time to get dark, but this is where it's such an incredibly powerful and high agency experience that does all these amazing things and it's also so dark and bad and scary. And it just points to the importance of us as designers, just to kind of put a bow on it, like knowing what's below the surface. Knowing that when we're designing these great experiences that help the person who bought the AI product go and do something, the data being collected has other impacts that may affect people well outside of that initial experience, but ultimately we are responsible for.

我想借此机会把视角拉远，因为考虑到我们讨论的一切、世界变化的速度以及现代设计实践所涉及的风险，这如何塑造了你作为设计领导者的表现方式，以及你思考管理、领导和投资组织的方式？

I think I want to take this opportunity to zoom all the way out then because given everything we're talking about and how quickly the world is changing and the stakes attached to the modern practice of design, even how is this shaping the way that you show up as a design leader and the way that you even think about managing and leading and investing into an org.

我在 Hacker Inc. 工作。我们帮助人们找工作。你进入我们的平台，展示你的技能。有很多人担心有人进来作弊，或者有意无意地做一些可能被视为影响结果的事情。一个 AI 副驾驶，本质上可以是你的个人监考员，只是在那里说，嘿，提醒你一下，当你切换标签查找语法时，这实际上会被负面记录，所以你最好别这么做。我们处理这个问题的方法是首先问：‘一个真正的监考员带来的出色体验是什么样的？当我开始时是什么样子？我想听到什么？我害怕什么？如果他们说了什么，我可能会怎么误解？如果他们需要干预怎么办？我可能会问哪些问题？他们能回答那个问题吗？’所以我们实际上创建了一个服务蓝图，描绘了一个出色的人本体验可能是什么样子。然后我们说，我们如何将其转化为软件？这就创造了一个新的框架：我们一直在构建软件即服务，现在几乎变成了先设计服务，然后说我们如何将其转化为软件。所以这是我们一直在做的一件事——我们做了很多抽象，这又回到了范围界定元素。通过这个过程，先考虑服务，然后说，好的，软件层是什么？AI 层是什么？我们已经有哪些上下文？我们如何以最无缝的方式收集它？这是我们经常做的事情。另一件事是意识到体验的设计不仅限于界面。提示词——软件本身的提示词——的配置方式或某个功能将极大地改变用户体验。理解不同类型或不同长度的上下文，或在特定时间共享的内容如何影响模型对你的响应，以及它如何容易地适应你并给你正确的选项。所有这些都影响用户体验。所以，我们一直在尽可能快地将设计转化为代码。不仅仅是因为设计师是否应该编码这个问题，而更多是因为模型本身现在已经成为体验的一部分。它实际上是体验的一方。

I work at Hacker Inc. And so we help people find jobs. You come into our platform and you demonstrate your skills. There's a lot of concern about people coming in and cheating, you know, or doing things intentionally or not that could be seen as influencing the results. An AI co-pilot that essentially could be your personal proctor, just there to say, hey, just so you know, when you switch tabs looking for syntax, it's actually going to be registered in this negative way, so you might not want to do it. And the way that we approached this problem was by first saying, 'What does a great experience look like with a real proctor? What does it look like when I get started? What do I want to hear? What am I afraid of? If they say something, how could I misinterpret it? What happens if they need to intervene? What types of questions might I ask? Would they be able to answer that question?' And so we actually created a service map of what an amazing human-centered experience could look like. And then we said, how do we translate this into software? And so it creates this new framing where we've been building software as a service and now it's almost like, well, design the service first and then say how do we translate this to software. So that's one thing that we've been doing is we've just been abstracting a lot and that's back into this scoping element. So going through that and actually thinking about the service first and then saying okay what is the software layer? What is the AI layer? What context do we already have? How can we collect it in the most seamless way? That's something that we're doing a lot of. Another thing is realizing that the design of that experience is not limited to the interface. The way that the prompt, the actual prompt of the software itself is configured or some feature is going to dramatically change that user experience. Understanding how different types or different lengths of context or things shared at certain times affects how the model responds to you and how easily it can adapt to you and give you the right options. That all affects the user experience. So, we've been trying to get designs into code as fast as possible. Not just because of this whole should designers code thing, but actually more because the model itself is now part of the experience. It's actually a party to the experience.

因此，我们不仅需要理解用户如何与我们创造的东西交互，还需要理解第三方如何影响他们的体验，以及我们如何为第三方设计，或者至少为用户更有效地引导它而设计。我想深入探讨你提到的那个点——你们如何更快地进入代码领域，这是我最近经常听到的一个主题。但我想了解，这如何改变了设计流程，以及设计师在你们团队中甚至与不同利益相关者协作的方式。这种变化带来了哪些差异？

And so we need to understand not just how does the user intersect with this thing we're creating, but how does this third party affect their experience and how do we design for them or at least design for the user's ability to direct it more effectively. I want to double click on the piece where you talked about how you're trying to get into code a little bit more quickly because that's a theme that I've been hearing. But I'd like to understand how that is changing the design process and how the way that designers in your or even collaborate with different stakeholders. What are some of the deltas that exist given that change?

这有点像“过去即是序章”，因为我不觉得我们离 15 年前有多远，那时还没有这些出色的原型工具，设计师常常需要用 HTML、CSS 和 JavaScript 做出一个足够好的版本。上了年纪的设计师都能告诉你，我懂一些基础的 CSS、HTML 和 JavaScript，因为这是我最有效地提前传达意图的方式。现在我们看到这些能力被抽象到了原型工具中。这并不完美——我就直说了。每个团队的运作方式都不同，我们的信念和理解程度也不一样，所以团队之间差异很大。但在那些有更多自由发挥空间（比如 AI 更原生地融入体验）的团队，或者我们对市场和设计对象有更强信念和理解的团队，我们正在快速推进到至少某种“活原型”阶段。我们使用的工具主要是 Figma Make 和 Lovable，这是大多数人用的两个，主要是因为方便。今天早上我和团队里一位设计师聊过，他正在做我们的 AI 数据产品。他到了某个点后说，把想法推到 Figma Make 里然后给工程师看“这是我设想的交互方式”，比尝试做原型要容易得多。没人觉得那会是最终版本，我们甚至不会进入开发模式，但它确实能更快地传达你脑子里的想法。这就是这些工具的用武之地。至于真正在代码库中工作，我们才刚刚开始尝试。很大程度上是因为，如果你不是从头开始搭建，单个团队、单个前端——比如设计工程师或与设计师紧密合作的前端工程师——可以走得很远。但由于我们在企业环境中工作，有非常严格的无障碍标准等，我们希望一起前进。所以现在我们正在做大量的运营基础工作，以便将更多设计工作迁移到实际开发工具中。我们的目标是到 2026 年底，所有设计师都能使用 Cursor。

There's a little bit of past is prologue because I don't know that we're that far off from where we were 15 years ago when we didn't have all these incredible prototyping tools and so like designers often had to get to a good enough version of something in HTML, CSS, and JavaScript. Like designers over a certain age can all tell you I have rudimentary CSS, HTML, and JavaScript because it was the most effective way for me to communicate my intent upfront. And so now we're seeing that be abstracted into these prototyping tools. It's imperfect. Like I'm just going to go ahead and say it. Every single team is operating differently. We have different levels of conviction and understanding. And so it's it really does look different from team to team. But on the teams where we have either a lot more of green space to play with like so AI is a little bit more native to the experience or where um we have a lot more conviction and understanding about the market and who we're we're designing for. Um yeah, we're moving pretty quickly into at least some sort of living prototype. And so the tools we're using there were are more Figma make um lovable. Those are the two that most people are using. Um, and and just because of the convenience factor. I was talking to a designer on my team this morning who's working on um something for our our like uh AI data product. And he got to a point where he was like, it's just so much easier for me to push this into Figma make and then show the engineer, hey, this is kind of how I'm thinking about this interaction than trying to prototype it. Nobody thinks that that's going to be the final version. Like we're not even going to bother going into dev mode. but it does create the ability to just convey what you have in your head a lot faster. And so, um, that's where those tools are fitting in. In terms of like actually working within the codebase, we're just starting to tiptoe into that. And, and a lot of the reason is that if you're not set up um to do that from scratch, individual teams, individual front end, you know, design engineers or front-end engineers working closely with a designer can can get pretty far. Um, but because we work in like an enterprise context with really strict accessibility standards and so on, we want to move forward together. And so right now we're doing a lot of the operational groundwork to let us be able to move more of design into our actual development tools. So we have a goal of all designers being in cursor by the end of 2026.

考虑到工作流程、工具以及协作方式的所有不确定性，当你在考虑想要雇佣哪种类型的设计师来与你一起踏上这段旅程时，这如何改变了你的优先考量？

Given all the uncertainty with workflows and tooling and how collaboration is changing, how does this shift what you're prioritizing when you're thinking about the types of designers that you want to hire to like, you know, set off on this journey with?

这是个很深的问题，因为真的，这又取决于具体情况，因为设计现在在很多不同方面都很重要。我喜欢那些来面试时眼睛发亮地告诉我“这是我用 vibe coding 做的，这是我在构建的东西”的人。现在最重要的技能是好奇心。好奇心之后，紧接着是“去搞定它”的态度。比如，我对某件事感到好奇，然后我去学了，然后卡住了。很好，我想聊这个。但你有自我驱动力说“我渴望学习新东西，然后我会去尝试弄明白”，这是最重要的技能组合。所以我告诉年轻设计师：去构建吧。去找个东西，然后去构建它。即使它不太管用也没关系，即使你永远不会把自己的个人凭证放进去也没关系，因为你只是在展示你如何开始塑造和打磨这个新事物的黏土，并开始形成自己的理解。这是很重要的一部分。我们经常谈论品味，但同样，这不仅仅是拥有观点。有很多审美很好的人，很难将审美转化到更广的范围，或者超越他们必须应用的具体问题。所以我们非常看重视觉设计，这是必须的——要有非常强的作品集，没有理由不做基础工作。但除此之外，我只想要那些已经开始形成自己关于什么有效、什么无效的语言的人。比如有人说，“嘿，我上周试了所有这些不同的产品，我想告诉你为什么这个比那个好。”这是另一个信号，表明这个人有很高的能动性，能够超越自己的视角或经验来看问题。这又回到了好奇心。我们肯定在向品牌倾斜，特别是那些思考更广泛体验的品牌设计师，因为品牌不止于网站或社交媒体上非常棒的页眉。品牌转化为信任层。比如，你在首次使用 AI 时，它的个性是什么？那就是品牌。Interaction 的 Poke 应用，我今年夏天早些时候开始用，它改变了我对聊天式交互的看法，因为它展示了一个模型的个性如何代表一家公司、公司的幽默感、公司的世界观。这不像某些工具只想钻到我的个人数据里——也许它确实想，但它很有趣，它会试图理解我，我喜欢和这样的人相处。我喜欢那种讽刺幽默，所以当然，我会给你钱。这是在大家都把价格降到每月 1 美元之前，所以我有点郁闷。但对我来说，那些思考每一个接触点、将产品作为社区一部分来使用的品牌设计师，就像被邀请加入一个氛围俱乐部——是的，我想给你我的数据，我想给你我的上下文，因为我信任你，因为我想成为这个群体的一部分。这就是品牌。所以我们需要品牌设计师超越眼前的画布去思考。我想最后一点是，那些能适应模糊性，并且乐于邀请他人进入模糊性的人。

That's a deep question because it really again it really depends because design is now important in a lot of different ways. I love it when people come to an interview and can tell me with their eyes lighting up like this is something I vibe coded. This is something I'm building. The most important skill is curiosity right now. Curiosity and then followed very quickly by go get them attitude, you know. So I was curious about something and then I went and learned it and then I got stuck. Cool. I want to have that conversation. But the fact that you have the self-direction to say, 'I am hungry to learn something new and then I'm going to go and try and figure it out.' That's the most important skill set. So I tell like younger designers, just go build. Just go and find something and go and build it. And it's okay if it doesn't really work. It's okay if you would never put your personal credentials inside of it because what you're doing is you're just showing how you can start to shape and mold this clay of this new thing and begin to develop your own understanding. So that's that's a big part of it. We talk a lot about taste, but again, it comes down to not just having an opinion. So there's a lot of people with great aesthetics who really struggle to translate beyond that aesthetic or beyond the immediate problems that they've had to apply it to. And so we are spiking on visual design. It's it's just a mustave like really strong portfolios. There's no excuse not to do the basics. But then beyond that, I just want people who have started to develop their own language of what's working and what's not. So someone who can say, 'Hey, I spent the last week just trying out all these different products and I want to tell you why this one worked better than that one.' That's another signal that this person has really high agency and can start to see beyond their own lens or their own experience. So it gets back to that curiosity piece. We're definitely leaning into brand, but particularly brand designers who are thinking about a wider experience because brand doesn't stop at the website or at, you know, a really awesome header, you know, on some social media site. Brand translates into that trust layer. Like what is the personality of AI when you first meet it during onboarding? That's brand. The poke app from interaction. I started using that earlier in the summer and it changed my mind about how we think about chatbased interactions because it showed how the personality of a model can represent a company, a company's humor, a company's sense of the world. Like this is not some some tool that's just trying to bury into my personal data. like maybe it is, but it's fun and it's going to try and understand me and like I like to be around people like that. I like the kind of sardonic humor, so heck yes, I'll give you my money. Um, this was before everybody was getting it down to $1 a month, so I'm a little bummed about that. But like to me, brand designers who are thinking about every single touch point, who are thinking about using the product as part of a community, it's part of like being invited into this vibe club that yes, I do want to give you my data. I do want to give you my context because I trust you because I want to be part of this group. That's brand. And so we need brand designers to be thinking beyond just the canvas in front of them. And I guess the last one I'd say is just just people who are comfortable with ambiguity and are comfortable inviting others into ambiguity.

但我认为这始终是设计不可或缺的一部分。每当利益相关者告诉我“这是我们应该做的，这是我的意见”时，我的下一步就是“太好了，我们去把它画出来吧。我们来做个工作坊。我要花 90 分钟和你一起，尽可能多地想出不同的方法来处理这个问题。”这样你的声音就摆在了桌面上。首先，人们会很快意识到，真正想出可行的概念并贯穿整个流程有多难。这让他们对你以及你的工作方式有了深刻的理解，同时也开始建立一种共同的语言和共同的视角，比如“我们到底要做什么？最终目标到底是什么？”这不是你的意见与我的意见之争，而是“我们服务的对象是谁，我们如何最好地服务他们？”因此，那些能够坦然邀请他人进入混乱、为混乱留出空间而不被淹没的设计师，现在成了一种超级能力。

But I feel like that's always been a necessary part of design. Like whenever I have stakeholders telling me this is what we should do, this is my opinion. My next step is great. Let's go sketch it out. Let's go do a workshop. I'm going to go spend 90 minutes with you and we are going to come up with as many different ideas for how to approach this as possible. So your voice is at the table. One, people realize pretty darn fast how hard it is to actually come up with viable concepts and play them through a journey. And so it gives them a deep understanding of you and how you're working, but it also starts to create that shared language and that shared view of like what are we actually trying to do? What is the actual endgame here? It's not your opinion versus mine. It's who is this person we're serving and how can we best serve them? And so designers who are comfortable of like inviting people into the mess and holding space for the mess and not drowning in it becomes just a superhero capability right now.

我很喜欢听你谈论好奇心，因为你显然也在付诸实践。你知道，你收集了所有这些文件夹和截图，而且作为设计领导者，你可能比许多实际制作界面的人负责更少的像素，但你仍然会说，“你知道吗，我要把玩所有这些，形成自己的观点，甚至磨练我的品味——不是在界面层面，而是在模型层面，以及这些更自然的语言交互看起来和感觉起来如何。”所以，我非常享受听到更多关于你如何思考、如何对待设计实践，以及这一切如何变化的内容，也很欣赏你的观点。非常感谢你今天来和我们分享，Emily。

I love hearing you talk about curiosity because it's so evident that you're putting it into practice too. You know, you have all of these folders and screenshots and as a design leader too, like you know, you might be responsible for fewer pixels than a lot of the people who are making these interfaces and still to say, you know what, I'm going to play with all of these and develop an opinion on them and even hone my taste not at the interface level, but at the model level and how these more natural language interactions look and what they feel like. And so I've just really enjoyed hearing more about how you think and approach the practice of design and how it's all changing and just appreciate your perspective. So thank you so much for coming on and sharing it with us today, Emily.

是啊，是啊。不，这真的很有趣。我想，邀请别人进入我自己的小混乱也挺有意思的。

Yeah. Yeah. No, this was really fun. It's fun to invite people into my own little mess, I guess.

在你走之前，我想花一分钟给你介绍一下我最喜欢的产品，因为我经常被问到我的工具栈是什么。Framer 是我用来建网站的工具。Genway 是我做研究用的。Granola 是我在评审时做笔记的工具。Jitter 是我用来给设计做动画的。Lovable 是我用代码实现想法的工具。Mobin 是我寻找设计灵感的地方。Paper 让我像创意人一样设计。而 Raycast 是我每一步的快捷方式。这些公司都是我精心挑选的，这样我才能全职做这些节目。所以，支持这个节目的首要方式就是去看看它们。你可以在 dive.com/partners 找到完整列表。

Before I let you go, I want to take just one minute to run you through my favorite products because I'm constantly asked what's in my stack. Framer is how I build websites. Genway is how I do research. Granola is how I take notes during crit. Jitter is how I animate my designs. Lovable is how I build my ideas in code. Mobin is how I find design inspiration. Paper is how I design like a creative. And Raycast is my shortcut every step of the way. Now, I've hand selected these companies so that I can do these episodes full-time. So, by far the number one way to support the show is to check them out. You can find the full list at dive.com/partners.