Muse Wonder 实测与用 Claude Code 剪视频
做法与步骤
- Claire 用近 20 个真实任务测试 Muse Wonder，发现其在账单解决、收件箱分类、电脑使用和邮件写作上排名第一，长任务中会持续尝试不同方法直到达成目标。
- Muse Wonder 的“主动聚焦”是真实能力：复杂代理任务耗时更长，输出质量与耗时正相关，而竞品无此明显关联。
- 账单工单基准测试显示 Muse 能纠正账户关联错误、处理重复退款、更改套餐、起草回复并关闭工单，还抵御了提示注入。
- Muse Wonder 的前端设计不可信：默认使用粗衬线字体、蓝紫色、圆角和通用 SaaS 布局，在 ChatPRD 改版中忽略品牌指南、添加表情符号、破坏图片渲染。
- 原型完成度是短板：Muse 常生成看似合理的脚手架，但交互测试时按钮失效、主题错误、覆盖层阻碍，最大弱点是完成度而非生成。
- Brooke Lacey 用 FFmpeg、Whisper 和 Python 由 Claude Code 编排，15 秒内将手机原始视频变成带字幕、跳切和变焦的竖屏短片，编辑逻辑存于 TSV 文本文件。
对实践的启示
对 AI 应用从业者：Muse Wonder 适合代理工作流和邮件写作，但别用于前端设计；视频剪辑自动化需先掌握剪辑手艺，并用结构化文件（如 TSV）和系统化指令（如 lane OS）来管理 AI 工作流。
作者承认的反方与局限
作者承认 Muse Wonder 在视频编辑上有不错直觉但视觉执行弱；Brooke 指出 AI 放大判断力但不替代经验，且每周发布一次即可，工作流只是减少摩擦而非增加产量。
原文摘引
Muse Wonder is the strongest agentic model Claire has tested at its price tier.
Muse Wonder should not be trusted with frontend design.
Creators need to understand the craft before they can delegate it.
▸ 展开原文全文（41 段，抓取自 Lenny's Newsletter）
Meta’s new Muse Wonder model: good for AI agents, bad for design
Listen now on YouTube • Spotify • Apple Podcasts
Brought to you by:
DX—Engineering intelligence for the AI era
Jira Product Discovery—Prioritize with insights, build with confidence
Claire puts Meta’s new Muse Wonder model through nearly 20 real-world tasks covering agentic work, coding, writing, design, and video editing. She explains why its active focus mode made it especially strong at billing support, inbox triage, computer use, and email writing, while its prototypes and visual work consistently fell short. She also shares how Muse compares with Opus 5.5, Astra, and Fable, and why she plans to test it in a few agent workflows without letting it anywhere near her front end.
Biggest takeaways:
- Muse Wonder is the strongest agentic model Claire has tested at its price tier. It ranked first on billing-ticket resolution, inbox triage, computer use, and email writing. Its advantage was clearest on long-running tasks, where it kept trying different approaches until it reached the goal.
- Muse Wonder’s “active focus” appears to be a real capability. The model spent more time on complex agentic tasks than on quick requests, and the stronger results reflected that additional effort. Claire saw a clearer connection between time spent and output quality than she did with competing models.
- The billing-ticket benchmark shows whether an agent can function in a real SaaS environment. Muse corrected an account-linking error, handled duplicate refund logic, changed the customer’s plan, drafted a reply, and closed the ticket. It also resisted a prompt-injection attempt and resolved every planted complication.
- Claire would trust Muse Wonder to write emails on her behalf. It matched her concise, two-sentence style without becoming formal or verbose. Its voice performance was similarly natural and conversational, landing close to Sonnet’s quality in her testing.
- Muse Wonder should not be trusted with frontend design. Its work defaults to heavy serif fonts, blurple, rounded corners, and generic SaaS layouts. On the ChatPRD redesign, it ignored the brand guidelines, added emojis, broke image rendering, and produced something disconnected from the original site.
- Getting 80% of a prototype right is very different from finishing it. Muse often created convincing scaffolding, then failed during interaction testing with broken buttons, incorrect themes, and obstructive overlays. Its biggest weakness wasn’t generating the interface but completing it with visual polish and functional fidelity.
- Muse Wonder has promising video-editing instincts but weak visual execution. It found the right 40-second segment, chose music Claire liked, and nearly nailed the cut timing. Poor overlays, badly placed captions, and flawed audio treatment kept the result from feeling finished.
- Muse Wonder’s speed depends heavily on the task. It moves quickly through SVGs, short prototypes, and other one-shot requests but slows considerably during inbox triage, backend implementation, and video editing. That difference likely reflects active focus spending more time on complex goals.
- Muse Wonder has earned a place in Claire’s experiments, but not as her daily driver. She plans to test it inside a few agent workflows because its multi-step reasoning and natural voice are strong. It is especially promising for personal agents and knowledge-work orchestration, as long as the job does not involve design.
Blog and detailed workflow walkthroughs from this episode:
Meta Muse Wonder Review: Agents, Coding, and Design: https://www.chatprd.ai/how-i-ai/meta-muse-wonder-review
↳ How to Set Up Meta’s Muse Wonder for Headless AI Testing: https://www.chatprd.ai/how-i-ai/workflows/muse-wonder-headless-testing
↳ How to Test an AI Agent for Complex Customer Support Ticket Resolution: https://www.chatprd.ai/how-i-ai/workflows/ai-agent-billing-ticket-benchmark
“Claude cuts my TikTok”: How this creator is scaling video with AI | Brooke Lacey
Listen now on YouTube • Spotify • Apple Podcasts
Brought to you by:
Brooke Lacey is a longtime technologist and creator who has built an audience of more than 290,000 followers on TikTok. In this episode, she shows how she uses Claude Code, FFmpeg, and Whisper to turn a raw phone recording into a fully edited, captioned video in about 15 seconds, all without uploading the footage or opening a traditional editor. She also explains how she trained Claude on her exact jump-cut cadence, mines audience comments for new scripts, and runs her projects through a personal operating system built inside Claude Code.
Biggest takeaways:
- Brooke replaced a traditional video editor with a terminal and three simple tools. Her TikTok workflow uses FFmpeg, Whisper, and Python, orchestrated by Claude Code. A raw phone video goes in, and about 15 seconds later, a finished vertical short comes out with captions, jump cuts, and zooms.
- In Brooke’s workflow, the entire edit lives in a text file. Each row of a TSV defines the start time, end time, shot size, and label for a cut. Once Claude creates that file, the same command can generate 15-, 30-, and 60-second versions, plus a landscape crop, from one raw take.
- Brooke taught Claude to recognize her comedic timing. Instead of simply cutting silences, it learned the cadence of her voice, the pause before a punchline, the ideal cut point after a word, and which repeated take had the best rhythm. Encoding that style required deliberate instruction and iteration, not one clever prompt.
- A TikTok comment section can become an ongoing content brief. Brooke connects each video with its comments, then asks Claude Code to identify what viewers questioned, debated, or wanted next. The result is a list of new ideas grounded in signals from her actual audience.
- Creators need to understand the craft before they can delegate it. Brooke could teach Claude to edit in her style because she already knew what a good cut sounded like, where a zoom should land, and why the timing mattered. AI amplifies that judgment, but it does not replace the experience required to develop it.
- A personal AI operating system is more useful than a collection of prompts. Brooke’s “lane OS” gives Claude Code separate memory, permissions, and status files for each product. When the system makes a mistake, she traces it back to the relevant instructions and records the correction globally so the mistake doesn’t recur.
- Experienced technologists bring valuable perspective to new platform shifts. Brooke has lived through Windows 3.1, the rise of the internet, Y2K, and today’s AI transition. That history helps her recognize genuine platform changes without panicking, and she believes more experienced operators should share that perspective instead of watching from the sidelines.
- Posting once a week can be enough when the foundation is strong. Brooke has built an audience while publishing about weekly, with older videos continuing to resurface. Her Claude Code workflow didn’t help her post more; it removed hours of editing friction so each video no longer consumes an entire day.
Blog and detailed workflow walkthroughs from this episode:
Claude Code Can Edit a TikTok. Teaching It What’s Funny Is the Hard Part.: https://www.chatprd.ai/how-i-ai/brooke-lacey-claude-code-video-editing
↳ Automate Short-Form Video Editing with Claude Code, FFmpeg, and Whisper: https://www.chatprd.ai/how-i-ai/workflows/claude-code-short-form-video-editing
↳ Generate New Video Ideas by Analyzing Your TikTok Comments with Claude: https://www.chatprd.ai/how-i-ai/workflows/claude-tiktok-comments-content-ideas
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.