四款个人AI代理实测与Figma PM的AI工作法
做法与步骤
- Claire用九项真实任务测试Grok Bot、Dot、Muse和Claude Duet，发现Grok Bot整体最强（93%），尤其适合工作场景，赢得Figma、收件箱分类和购物任务。
- Muse在个人生活任务上表现最佳，快速生成短视频和家庭报纸，速度优势让Claire将部分快速问答从ChatGPT迁移到Muse。
- 速度是产品体验核心：Muse在九项任务中四次率先完成，而Claude Duet始终最慢，代理消失几分钟会让用户失去上下文。
- Claude Duet请求许可过于频繁（基准测试中四次，其他仅一次），让用户感觉需要监督而非信任。
- 代理能力受限于可访问工具：Codex审计发现各代理连接器和技能不等，Muse缺少Slack和GitHub导致无法完成部分工作任务。
- Figma的PM OS基于31项AI技能，从PM常忽略的任务入手（如启动前联系支持、记录会议），建立最低可接受标准；通过“what’s up”技能每天六次管理Slack，归档约10个频道。
对实践的启示
从业者应放弃寻找全能代理，按工作与生活场景分别选型；同时可借鉴Figma PM OS，用共享技能库和自动化代理处理常规任务，并注意代理能力受工具连接限制，需确保公平比较。
作者承认的反方与局限
作者承认代理能力受工具访问限制，比较可能不完全公平；Claude Duet的谨慎请求许可虽出于好意，但影响体验。
原文摘引
Grok Bot is the strongest all-around personal agent in this benchmark, especially for work.
The useful takeaway is that different agents belong in different parts of life.
When an AI conversation goes bad, starting over often beats arguing with it.
▸ 展开原文全文（42 段，抓取自 Lenny's Newsletter）
Grok vs. Dot vs. Muse vs. Claude Duet: which one crushes the competition?
Listen now on YouTube • Spotify • Apple Podcasts
Brought to you by:
WorkOS—Make your app enterprise-ready, with SSO, SCIM, RBAC, and more
Firecrawl—Give your AI agents web data and beyond
In this solo episode, Claire puts Grok Bot, Dot, Meta’s Muse, and Claude Duet through nine real-world tasks to find out which personal agents are actually worth using. She tests them on everything from Slack replies, inbox triage, and coding to video editing, Figma thumbnails, family projects, and shopping. The results point to a clear split.
Biggest takeaways:
- Grok Bot is the strongest all-around personal agent in this benchmark, especially for work. It scored 93% overall and won the Figma, inbox triage, and shopping tasks. For someone choosing one agent for professional workflows, Claire recommends starting here.
- Muse is the best agent for personal life. It created the strongest short-form video and family newspaper with almost no prompting and answered a physics question too quickly to time. That speed has led Claire to move more quick Q&A conversations from ChatGPT to Muse.
- Speed is central to the product experience. Muse finished first on four of the nine tasks, while Claude Duet was consistently the slowest. When an agent disappears for several minutes, users lose context and stop wanting to use it.
- Claude Duet asks for permission too often. It requested approval four times during the benchmark, compared with once for every other agent. That caution may be well-intentioned, but it makes Duet feel like something the user must supervise rather than trust.
- An agent is only as capable as the tools it can access. A Codex audit found that Claire’s agents did not have equivalent connectors and skills. Muse lacked Slack and GitHub, which prevented it from completing several work tasks, highlighting the importance of equal setups for a fair comparison.
- Grok Bot is the agent Claire trusts most with Figma. Dot and Claude Duet mostly filled existing templates, while Grok Bot proposed fresh headlines and layouts that looked like real YouTube thumbnails. It has consistently outperformed the others across Claire’s Figma tests.
- Claude Duet and Dot still showed the strongest coding judgment. Dot came closest to the intended Sentry fix, while Duet found a second bug and avoided duplicating work Dot had already completed. Claire would still choose Claude or Codex for serious coding, just not through Duet’s personal-agent interface.
- Dot showed the strongest personal context in one important moment. It was the only agent that asked whether Claire’s Lovable talk was paid and remote, which felt like genuine chief-of-staff judgment. Its poor performance on creative tasks, however, kept it well behind Grok Bot and Muse overall.
- The useful takeaway is that different agents belong in different parts of life. Claire’s current split is Grok Bot for work and Muse for personal tasks. Choosing agents by job is more effective than expecting one product to handle everything well.
Blog and detailed workflow walkthroughs from this episode:
Grok vs. Dot vs. Muse vs. Claude Duet: Nine Tasks, Four Agents, No Universal Winner: https://www.chatprd.ai/how-i-ai/grok-vs-dot-vs-muse-vs-claude-duet-nine-tasks-four-agents-no-universal-winner
↳ Generate Novel YouTube Thumbnail Ideas with Grok and Figma: https://www.chatprd.ai/how-i-ai/workflows/generate-novel-youtube-thumbnail-ideas-with-grok-and-figma
↳ How to Create a Branded Product Launch Video with Claude Duet: https://www.chatprd.ai/how-i-ai/workflows/how-to-create-a-branded-product-launch-video-with-claude-duet
↳ Create a Personalized Family Newspaper with Meta’s Muse: https://www.chatprd.ai/how-i-ai/workflows/create-a-personalized-family-newspaper-with-meta-s-muse
How Figma’s PMs actually use AI every day | Ezra Mechaber (PM, Figma)
Listen now on YouTube • Spotify • Apple Podcasts
Brought to you by:
Optimizely—Your AI agent orchestration platform for marketing and digital teams
Jira Product Discovery—Prioritize with insights, build with confidence
Ezra Mechaber is a growth product manager at Figma and one of the creators of PM OS, a shared repository of 31 AI skills used across the company’s product organization. In this episode, he demonstrates how Figma’s PMs use it to manage Slack, gather organizational context, create artifacts, and run more thorough bug bashes with Codex. He also walks through AutoXP, an AI growth PM that runs experiments with a single human checkpoint, and a real-time Jev assistant that improves writing as he types.
Biggest takeaways:
- The best place to start with AI is the work PMs routinely skip. Ezra designed PM OS around tasks that fall through the cracks, such as involving support before a launch or documenting meetings. His goal is to establish a minimum acceptable standard for every PM workflow.
- Context-gathering skills are the foundation of useful AI at work. At a company with roughly 2,000 employees, even identifying the right collaborators can be difficult. One PM OS skill maps a team and automatically adds the relevant designer, engineering manager, and team leaders as reviewers.
- A simple “what’s up” skill can make Slack manageable. Ezra runs it about six times a day to gather relevant threads, projects, open decisions, and people awaiting a response. He archives roughly 10 channels daily because the skill preserves what actually matters.
- A basic Git repository can support an entire PM organization. PM OS consists of an AGENTS.md file, reusable skills, and a few command hooks. An automatic version check keeps 60 to 70 PMs aligned, while private subrepositories give each person room to customize.
- Successful AI adoption starts with familiar tools, then evolves with users. PM OS began as a terminal interface in Claude Code before gaining a local web interface that was easier to read. Ezra now works primarily in Codex, partly because browser use has become reliable enough for his workflow.
- Agents can be more diligent bug bashers than humans. Ezra gives his agent a test account and a FigJam board full of scenarios, then has it document its findings. It has caught edge cases ranging from a text box filled with 1,000 emojis to grammar that changed based on a team’s name.
- In growth, wasted traffic is the biggest sin. AutoXP uses three agents to identify funnel drop-off, propose new copy, and prepare the implementation. A human reviews the experiment once before launch, after which the system analyzes results and begins the next iteration automatically.
- Fast, inexpensive models can provide feedback that feels continuous. Ezra’s Jev prototype reviews writing whenever the user pauses, flags placeholders and unsupported claims, and finds relevant interview quotes from a research library. The speed makes it feel less like running an AI task and more like working beside an attentive editor.
- When an AI conversation goes bad, starting over often beats arguing with it. Ezra treats poor output as a sign of unclear instructions or contaminated context. He opens a fresh conversation, restates the goal, and approaches the model with high expectations: “Buddy, I know you can cook.”
Blog and detailed workflow walkthroughs from this episode:
Inside Figma’s PM OS: 31 AI Skills, Better Bug Bashes, and an Automated Growth Loop: https://www.chatprd.ai/how-i-ai/inside-figma-s-pm-os-31-ai-skills-better-bug-bashes-and-an-automated-growth-loop
↳ Automate Software Bug Bashing with a Browser Agent: https://www.chatprd.ai/how-i-ai/workflows/automate-software-bug-bashing-with-a-browser-agent
↳ Build an Automated Growth Experiment Loop with Multiple AI Agents: https://www.chatprd.ai/how-i-ai/workflows/build-an-automated-growth-experiment-loop-with-multiple-ai-agents
↳ Build a Personal PM Operating System with a Shared Knowledge Repo: https://www.chatprd.ai/how-i-ai/workflows/build-a-personal-pm-operatin