Meta设计Muse：让AI像人一样对话
做法与步骤
- 设计原则：以“如果和真人对话会怎样”为出发点，去掉传统消息气泡，将代理置于屏幕顶部而非标准聊天窗口内。
- 渐进式披露：同一活动分三层呈现——休闲用户看表情符号，需要细节的看任务列表，工程师看原始日志，各取所需不增杂乱。
- 可见思维链常是延迟的变通：Muse足够快，能安静运行，无需表演思考过程，直接交付结果。
- 生成式UI：结构固定、内容变化。头像、信息流和想法卡片是固定表面，动态元素是状态表情、个性化建议和微动画。
- Instagram收藏是消费级AI被忽视的数据源：连接多年收藏，让Muse整理并生成SkyMall风格购物推荐，无需技术设置。
- 设计非确定性系统需改变对工艺的认知：Muse因人而异，无法保证像素完美，以有用性而非视觉精度评判结果是解放性转变。
对实践的启示
做AI应用时，应优先考虑对话式交互而非传统UI，用渐进式披露平衡不同用户需求，固定界面结构而让内容动态化，并重视个人数据（如收藏）的再利用。
原文摘引
How would this work if you were talking to another human?
Visible chain-of-thought is often a workaround for latency, not a valuable feature.
Generative UI works best when the structure stays fixed and the content changes.
▸ 展开原文全文（23 段，抓取自 Lenny's Newsletter）
How Meta designed Muse: an AI agent people actually want to talk to | Alex Cornell (Design Director)
Listen now on YouTube • Spotify • Apple Podcasts
Brought to you by:
WorkOS—Make your app enterprise-ready, with SSO, SCIM, RBAC, and more
Alex Cornell is a design director at Meta, where he leads design for Muse. In this episode, he explains how progressive disclosure makes Muse approachable for casual users without limiting technical ones, why its main chat is designed as a single persistent relationship, and how generative UI can feel personal without constantly changing the interface. He also shares how he uses Muse to turn years of Instagram saves into a shopping catalog, analyze his golf and chess games, and transform a 20-year-old blog into a daily podcast.
Biggest takeaways:
- AI agents should be designed for conversations with people, not interactions with software. Alex’s guiding question for Muse was “How would this work if you were talking to another human?” That principle shaped everything from removing traditional message bubbles to placing the agent at the top of the screen instead of inside a standard chat window.
- Progressive disclosure is one of the most useful principles in AI product design. Muse represents the same activity at three levels: a simple emoji for casual users, a task list for anyone who wants more detail, and raw logs for engineers. Each person gets the transparency they need without adding clutter for everyone else.
- Visible chain-of-thought is often a workaround for latency, not a valuable feature. Many AI products display reasoning steps to reassure users while a slow model works. Muse is fast enough to operate quietly, showing that agents do not need to perform their thinking when they can simply deliver the result.
- Generative UI works best when the structure stays fixed and the content changes. Muse’s avatar, feed items, and idea cards are deliberately designed surfaces. The dynamic elements are the agent’s status emoji, personalized suggestions, and small animations, allowing the experience to feel alive without rebuilding the interface every session.
- Instagram saves may be one of the most overlooked data sources for consumer AI. Alex connected years of saved posts to Muse, asked it to organize them, and turned the results into shopping recommendations styled like a SkyMall catalog. It gave forgotten personal data a practical new use without requiring technical setup.
- Designing non-deterministic systems requires a different relationship with craft. Alex has spent his career controlling exactly how interfaces appeared, but Muse looks different for every person and cannot guarantee pixel-perfect outputs. Learning to judge the results by usefulness rather than visual precision has been one of the most liberating shifts in his career.
- Muse’s single persistent chat is a meaningful product bet. Instead of dividing every interaction into separate threads, it treats the main conversation as an ongoing relationship with the agent. Side chats handle scoped tasks, but the central thread preserves continuous context in a way Alex believes the industry has not fully explored.
- Polite prompting may be an effective strategy, not merely good manners. Alex responds to frustrating results with calm, specific feedback rather than anger. He’s also seen agents change their responses when users express frustration more sharply. His conclusion is that clear, respectful direction may produce better results.
Blog and detailed workflow walkthroughs from this episode:
Alex Cornell’s 3 Meta Muse Workflows for Shopping, Hobbies, and Archives: https://www.chatprd.ai/how-i-ai/alex-cornell-meta-muse-workflows
↳ How to Create a Personalized Shopping Catalog from Your Instagram Saves Using AI: https://www.chatprd.ai/how-i-ai/workflows/meta-muse-instagram-shopping-catalog
↳ How to Use AI for Hobby Improvement and Data Analysis for Golf & Chess: https://www.chatprd.ai/how-i-ai/workflows/meta-muse-golf-chess-analysis
↳ How to Revive an Old Blog or Digital Archive as a Personal AI Podcast: https://www.chatprd.ai/how-i-ai/workflows/meta-muse-blog-archive-podcast
If you’re enjoying these episodes, reply and let me know what you’d love to learn more about: AI workflows, hiring, growth, product strategy—anything.
Catch you next week,
Lenny
P.S. Want every new episode delivered the moment it drops? Hit “Follow” on your favorite podcast app.