
AI Engineer | CTO & Co-Founder at Snapbar | Forbes 30u30 - Seattle, '23 | Co-Founder of @DubHacks (prev.)
Seattle, WA
Joined May 2008
- Tweets457
- Following376
- Followers1.3K
- Likes2.6K
Pinned Tweet
Replying to @awilkinson
The most powerful workflow I use is Claude Code (or Cursor) paired with the Playwright MCP (github.com/microsoft/playwri…) which allows coding agents to navigate a browser, take screenshots, read network/console logs, etc. It's the missing link to enable autonomous front-end iteration (without your input).
It is truly magical in dramatically increasing the quality of the output, especially when 'designing' of building UIs. These LLM's are only as good as the context they have, and what they can 'see'. Playwright allows them to see their UI work, and iterate against it.
Patrick Ellis retweeted
Been building an open source Claude Tag on @flueai with self-hosting on Cloudflare.
One-click to deploy and includes an admin where you can:
- set any LLM provider & model
- define multiple agents and scope each to certain Slack channels
- add skills, MCPs, and custom instructions per agent
Striving to be Claude Tag without the lock-in.
If anyone is interested in trying it out hmu as it’s getting close to the finish line!
Beads X GStack
+1 for Yegge's Beads / Gas Town. @garrytan I've been scaling our startup with Claude Code / etc since it's release Feb '25, and GStack + Beads has been the largest unlock I've come across, alongside Playwright MCP as verifier.
Making a YouTube vid on it soon - thank you! 🙏🙏
Replying to @nrmehta
I think you have to make your own tools and use them to speed up
You need more personalization, more of your thoughts and ideas in markdown supporting what YOU want
Figure out the decision points and reduce the need for you to personally intervene
Find leverage
Turns out, all this time investing in my Second Brain over the past 6yrs may have not been indulgent. Crazy high leverage with personal assistants (e.g. OpenClaw). Which will only compound as the models, agents, and tooling gets better.
The one thing foundations labs don’t have is context about you - your goal, business, life, etc. Collecting and organizing that info is key to executive assistant tasks, and other agentic tasks.
Best DX for running Claude Code (etc) in a local sandbox (so you can safely skip the permission checks with the --dangerously-skip-permissions flag).
Love the idea of CC's /sandbox command, but have sadly found it doesn't fully respect my settings.local.json permissions.
For devs asking “how do I run coding agents without breaking my machine?”
Docker Sandboxes are now available.
They use isolated microVMs so agents can install packages, run Docker, and modify configs - without touching your host system.
Read more → bit.ly/4teSVgL
Patrick Ellis retweeted
I'm Boris and I created Claude Code. Lots of people have asked how I use Claude Code, so I wanted to show off my setup a bit.
My setup might be surprisingly vanilla! Claude Code works great out of the box, so I personally don't customize it much. There is no one correct way to use Claude Code: we intentionally build it in a way that you can use it, customize it, and hack it however you like. Each person on the Claude Code team uses it very differently.
So, here goes.
Patrick Ellis retweeted
Don't think of LLMs as entities but as simulators. For example, when exploring a topic, don't ask:
"What do you think about xyz"?
There is no "you". Next time try:
"What would be a good group of people to explore xyz? What would they say?"
The LLM can channel/simulate many perspectives but it hasn't "thought about" xyz for a while and over time and formed its own opinions in the way we're used to. If you force it via the use of "you", it will give you something by adopting a personality embedding vector implied by the statistics of its finetuning data and then simulate that. It's fine to do, but there is a lot less mystique to it than I find people naively attribute to "asking an AI".
Opus 4.5 turns Claude Code into an excellent front-end designer IF you take off its "blindfold".
All the vision improvements in Opus 4.5 do no good if Claude Code has no ability to visually check its work. The solution: Playwright MCP → browser control + screenshots (aka “Claude Code's eyes”).
Less literary, Claude Code must also be able to understand your vision for good UI design. Add a Skill (or dedicated Subagent) personalized to direct Claude Code's style and use of Playwright. Start with Anthropic's frontend-design skill template and customize for design best practices (links below).
I do a deep dive on exactly how we're using this workflow in this YouTube video (with 229K+ views): youtube.com/watch?v=xOO8Wt_i…
I outline the exact customizations to make to your CLAUDE.md + Subagent (or Skills) configs in the video. You can easily modify Anthropic's default front-end Skill (github.com/anthropics/skills…) using the general design principles I outlined in this config example (github.com/OneRedOak/claude-…), along with your brand's specific customizations.
Patrick Ellis retweeted
The main lesson of the past few weeks is that the Big Four US labs all seem to have figured out a path forward in continuing the exponential pace of LLM improvement, at least in the near future.
As a result, agents continue to advance in coding & in office tasks like PowerPoint
Patrick Ellis retweeted
New on the Anthropic Engineering Blog: Most developers have heard of prompt engineering. But to get the most out of AI agents, you need context engineering.
We explain how it works: anthropic.com/engineering/ef…
Patrick Ellis retweeted
Shout out to @PatOakEllis I have found his videos to be some of the most helpful content for learning "context engineering". This one for example was a game changer for my workflows, giving your agent "eyes" youtu.be/xOO8Wt_i72s?si=ywDA…
Patrick Ellis retweeted
Announcing our public preview of Chrome DevTools MCP! Experience the full power of DevTools in your AI coding agent → goo.gle/4pDE6Tk
With Chrome DevTools MCP, your AI agent can run performance traces, inspect the DOM, & perform real-time debugging of your web pages.
Patrick Ellis retweeted
Claude Code + Playwright MCP == 更强的前端设计能力 —— 来自 @PatOakEllis 的视频教程
核心问题:为什么 AI 生成的设计总是千篇一律?
Patrick 指出,很多人在使用 cc 或其他 AI 编码工具(如 Cursor)时,期待生成漂亮、现代的 UI 设计,结果却总是得到类似 “紫色 ShadCN” 风格的普通界面,像 Twitter 上随处可见的模板化设计。而且,调整设计到像素级完美几乎不可能,感觉像在和 AI “绕圈子”。问题不在 AI 本身,而在于我们给它的环境限制了它的能力。AI 就像一个“蒙着眼睛的博士”,只能通过代码“摸索”设计,无法“看到”自己的成果。
解决之道:用 Playwright MCP 给 cc “开眼”
Patrick 分享了一个关键突破:通过 Playwright MCP,让cc 拥有“视觉”能力。Playwright 是一个由微软开发的浏览器自动化框架,主要用于网页测试,但它的 MCP 可以让Claude控制浏览器、截图、检查页面,并根据视觉反馈迭代优化设计。这相当于解锁了 cc 的多模态能力(不仅限于代码,还能处理图像),让它能像人类设计师一样“看到”并改进 UI。
三大核心理念
1. 编排层(Orchestration Layer)
要让 cc 发挥最大潜力,需要给它提供完整的上下文、工具和验证机制。上下文包括清晰的提示词和文档,工具如 Playwright,验证机制则是 UI 原型或风格指南等明确的标准。这样,AI 才能清楚目标是什么、用什么工具、如何判断结果好坏。
2. 迭代智能体循环(Iterative Agentic Loop)
设计不是一次生成就完美,而是需要反复调整。Playwright 让 cc 能查看截图,比较当前设计与目标(如风格指南)的差距,自动修正并再次检查。这个循环让 AI 可以独立工作更长时间(甚至半小时以上),无需人工反复干预,最终产出更接近预期的结果。
3. 挖掘模型的视觉智能
Claude 的训练数据包含大量文本、代码和图像,但普通使用方式只调用了它的代码能力,忽略了视觉理解的部分。通过 Playwright 截图,cc 能利用其视觉智能,分析 UI 元素、布局、配色等,像一个真正的设计师一样思考。
Playwright MCP 的强大功能
· 自动截图:让 cc “看”到页面,检查 UI 是否符合预期。
· 读取浏览器日志:识别控制台或网络错误,自动修复。
· 模拟设备:测试不同设备(如 ·iPhone)或屏幕尺寸的响应式设计。
· 浏览器导航:模拟用户操作(如点击、填表单),复现特定状态。
· 抓取参考页面:从目标网站截图或提取灵感。
· 自动化测试:包括端到端测试和无障碍性检查。
七个实用工作流
1. 迭代前端设计:通过截图和日志,cc 可以持续优化 UI,产出更精美设计。
2. 自动修复 UI 错误:发现并修复控制台中的明显问题。
3. 模拟用户操作:根据用户描述(如“点击某按钮后出错”),复现问题并解决。
4. 抓取参考设计:从指定网站获取灵感,生成类似风格的 UI。
5. 端到端测试:自动化测试整个应用流程。
6. 无障碍性检查:确保设计符合可访问性标准。
7. 响应式测试:快速检查桌面、平板、手机等不同设备的显示效果。
配置与实战
1. 安装 Playwright MCP
安装简单,只需在 cc 中运行一行命令即可(视频中展示具体代码)。可以配置浏览器类型(如 Chromium)、设备模拟(如 iPhone 15)、是否无头模式(默认有界面,方便观察)。
2. 优化 claude. md 文件
claude. md 是 cc 的“记忆”文件,记录每次会话的默认设置。Patrick 在其中添加了:
· 设计原则文档(基于 Gemini 深度研究整理的审美标准)。
· 风格指南和 UI 原型,作为验证标准。
· 自动化规则:每次前端改动后,cc 自动打开浏览器、截图、检查错误,并根据提示中的验收标准调整。
3. 自定义子智能体(Sub-Agent)
Patrick创建了一个“设计审查”子智能体,专门审查 UI 变化(如 PR 或新提交)。它会:
· 使用 Playwright 检查页面。
· 参考设计原则和风格指南。
· 提供详细报告,指出优点、高优先级问题(如性能瓶颈)和改进建议。
· 可由其他子智能体调用,形成复杂的工作流网络。
4. 并行开发与 Git Worktrees
为利用 AI 的随机性(每次生成结果略有不同),Patrick 使用 Git Worktrees 创建多个分支,同时运行多个 cc 实例,生成不同设计版本,再挑选最佳结果。这种“并行实验”能显著提高产出质量。
实用技巧
· 学习 Anthropic 的官方示例:参考他们的 GitHub 和文档,了解如何结构化 claude. md、子智能体和命令。
· 善用视觉上下文:在提示词中加入截图、低保真草图、风格指南或参考网站链接,激发 cc 的视觉智能。
· 封装专家知识:将团队中优秀设计师或工程师的经验提炼成子代理或命令,供全员复用。
· 深研辅助:用 Gemini 等工具研究设计最佳实践,整理成简洁文档,供 cc 参考。
Patrick Ellis retweeted
Been deep in Claude Code for months now, so this from @PatOakEllis hit home: Practical workflows from an AI-native startup.
The Design Review workflow is kinda cool: Automated feedback on frontend PRs for UI and UX consistency and accessibility before merge. Less guesswork, more signal.
Under the hood, it drives a real browser with Playwright MCP and pairs it with Claude Code agents. It looks at the rendered UI, checks against standards, and catches visual issues early.
This turns design QA into a repeatable step teams can share, not a last minute ping. Handy when you are growing a design system or shipping fast.
Patrick Ellis has solid tutorials and demos on YouTube. Worth a look: github.com/OneRedOak/claude-…
Patrick Ellis retweeted
An incredible video packed with very valuable information on how to leverage the power of Claude, MCPs, Agents, Playwright in order to 10x your frontend development with AI!
Thanks @PatOakEllis for taking the time and efforts to teach us those productivity hacks! 🙏
youtu.be/xOO8Wt_i72s?si=ch1R…
#claude #claudecode #AiAgents #mcp #vscode #productivity #UIDesign #FrontendDev
















