Product Designer & Builder ｜Everyday share my design, UI, tools, AI & weird ideas and thoughts 产品设计师 & Builder，每日分享我的设计、工具、AI 、灵感和思考
United State
Joined August 2025
- Tweets616
- Following953
- Followers76
- Likes639
time to draw up an army to start the crusade🗡️
teapawt.org/Invouxly/status/211617…
well, testing time, now I see a sturdy soldier blocking the enemy’s arrows with a shield
teapawt.org/Invouxly/status/211617…
Inspired by @oiharshit 's loading designs and the emotional design idea behind them. When you can see something working for you the whole time, the wait feels shorter. The cat is my own take on that.
teapawt.org/oiharshit/status/21146…
对于 Macbook 和 Windows 笔记本双持党来说（我就是），习惯了 MacBook 的三指拖放，回到 Windows 触控板上手指就跟吃了屎一样难受。分享一个可以在 Windows 上实现三指拖放的项目，这下体验真的好得多
github.com/ClementGre/ThreeF…
took it a step further. built the game into my Codex desktop app, so my desktop now lives in the ocean. get eaten and you respawn right back on the desktop. finally something to do while waiting on the agent during vibe coding. 🎮🐠
teapawt.org/Invouxly/status/211507…
Just used the brand new GPT-6.1 Terra to make an ASCII "big fish eat little fish" game! Ultra tier, and the model wrote the code.
It’s all rendered in text characters. You start out as a tiny minnow and eat your way up to swallowing great white sharks, eventually becoming the king of the ocean🦈👑
Live: aenvo.github.io/ascii-big-fi…
Github: github.com/Aenvo/ascii-big-f…
Just used the brand new GPT-6.1 Terra to make an ASCII "big fish eat little fish" game! Ultra tier, and the model wrote the code.
It’s all rendered in text characters. You start out as a tiny minnow and eat your way up to swallowing great white sharks, eventually becoming the king of the ocean🦈👑
Live: aenvo.github.io/ascii-big-fi…
Github: github.com/Aenvo/ascii-big-f…
On the same page. Nobody's happy with AI's first version unless they're new to the craft. the polish runs on feel. Even iconic logos aren't perfect by the math, the thinking and resonance matter more. The idea doesn't have to be brand new, but what you add has to be yours, and you credit the source.
I've been sharing a lot of interactions over the past week, and a few people have asked how I make them.
A few things about my process:
1. Start with an idea worth sharing. I mostly don't involve AI at this stage because I want the thinking to be mine.
2. Figure out how to communicate that idea through motion. The animation should add something, not just make it look cool for posting.
3. Design the actual elements manually in Figma, Paper, or whatever tool works best. You're a designer.
4. Use AI to bring the interaction to life. If you know how to code, you can give your agent more specific instructions from the start. I use different models for different stages, but I rarely get something I'm happy with in the first three attempts.
5. Spend most of the time refining. Iteration is the key. Timing, easing, transitions, typography, all the little details that make an interaction feel right.
6. Once it feels good, render and share.
The important distinction is knowing what you want before asking AI to build it. The tools make execution faster, but they don't replace a clear point of view, your experience and expertise.
Haven't tried @elayadesigns's ai-design-skills yet, but these landing pages look really good. Saving it for later. Thanks for open sourcing it
I designed all of these landing pages in the last two months, using @higgsfield , Claude, and Cursor with my custom AI designer skill.
The skill is open source: github.com/elayadesign/ai-de…
码住了，这两位也是我非常喜欢和认可的设计师，可以多多关注
按钮是最能影响设计语言的组件之一，把它做好了边际效益很高。
可以看看@ricouii 推荐的这个仓库，直接抄作业 evilbuttons.com/
Saw @zzzzshawn's excellent small plays, then @awwbhi2 made a Pokémon battle in that style and shared how. I followed his workflow in Codex: measure the frames, study the moves, build in three.js, render frame by frame.
Dragon Ball Z: Goku vs Frieza on Namek. Thank you both!!!
一份 SEO 自查清单分享，使用 AI 做网站可以留意一下
很多用 AI 几天搓出来的网站在 Google 搜不到，根本不是内容差，而是上线前漏了这 7 项。
Vibe Coding 确实快，但大部分人写完前端往 Vercel 一挂就以为万事大吉，
一个月过去后台展现量还是鸭蛋。
我给你整理了一份出海网站上线前必须打勾的 SEO 救命清单：
1. 检查是不是纯客户端渲染：必须配置 SSR（服务端渲染），确保爬虫抓到的是完整 HTML；
2. 生成动态 Sitemap.xml 并主动提交到 Google Search Console；
3. 检查 robots.txt：千万别手滑把 Googlebot 给拒之门外；
4. 全站全局搜一下 noindex 标签，很多脚手架测试环境默认带这个，上线前必须干掉；
5. 清除多重重定向死链，把 404 页面做优雅降级。
这 7 项自查加起来不到半小时，但决定了你几个月的心血能不能被搜索引擎正常索引。
建议做独立开发的朋友先收藏，每次新站上线前照着逐项对一遍。
读完 @JasonPang0720 这篇，第一反应是想到通用类型 Skill、特定工作流 Skill，还有 AGENTS.md 里的各种约束
搭 Agent 工作流时，大家容易只想着该加什么规则，很少想这条规则该在多大范围里生效（对规则命中的考虑）。而任务一旦复杂，全局约束和特定约束不分开，这里就容易出问题，常见场景就是，本来是为了让它更靠谱的要求，反而把 Agent 绑住了
这跟 OPRD 论文（arxiv.org/abs/2610.21496）里的 Multi-Scope Reminders（多作用域提醒）有点像，它把提醒分成两类：
1. Global Reminder 贯穿整条任务轨迹；
2. Turn-local Reminder 只管某次工具反馈之后的下一步，免得过时的纠正干扰后面的动作。
在这篇论文的过程诊断里，同样 200 个任务，错误事件从原模型的 446 次降到训练后的 222 次。注意这是错误事件数，不是失败任务数。这说明至少在这个训练里，这种分法是有用的。
OpenAI 的《Rethinking skills and prompts for GPT-6 Astra》（developers.openai.com/blog/r…）里也有个很典型的例子：别让 Agent 每次改代码前都去读 architecture.md、database.md、deployment.md，碰到服务边界、数据库结构、部署的时候，再分别读对应的那份。Skill 的触发描述写得太宽，会在不相关的任务里被加载。
所以我现在按作用域放规则，分成 5 条：
1. 全局规则：只放长期有效的行为习惯和安全边界，每个任务都要遵守，尽量简短
2. AGENTS.md：分仓库级和目录级，只管对应项目和模块的约束
3. Skill：只在满足条件时触发某个工作流，有些 skill 我会关掉隐式触发，只在我点名时才用
4. 当前任务：只写这一次的要求，不往大了说，也不管任务以外的事
5. 局部反馈：当场纠偏，用完就丢，不留到后面的任务里
特别是这两种反方向的错配要小心。一种是把一次性的纠正写成全局规则，之后每个任务都被它影响。另一种是把关键的安全约束放进不一定会触发的 Skill，Skill 没被调用，安全约束也跟着没了
要明确每条规则管哪儿、什么时候触发、管多久，条数多少倒在其次
当然，这只是从 OPRD 延伸出来的运行时规则组织思路，跟论文的训练方法不是一回事儿。Jason 在文末也提醒了，光在提示词里加类似的话，拿不到训练带来的那部分收益
你们是怎么做的呢？
