yrzhe @foolsphilo

看到一个好玩的东西叫 mindwalk，试了一下，一个把 agent 会话回放成仓库 3D 地图的小工具。 碰过的文件亮起来，没碰的黑着，还能切文件树和 3D 地形两种视图。不过当生产力工具还谈不上，感觉更多是好玩。 选了一个 Session 然后拉完整张图基本是黑的，一万两千多个文件，它真正碰过的就两百来个，全挤在一个角落，剩下九成八全程没点开过。那一小片里还有二十九个文件被来回改了三四遍以上，近一半读取在重读同一批。 所以所谓 agent 理解你的代码库，多数时候就是盯着那一小撮反复磨，整个仓库九成八它根本没进去。不过也可能是我的 agent 在执行特定任务，所以没有特别去看其他的地方。但总之用处感觉不是很大。。

今天试了下 ChatGPT 网页端的 Work ，给它装了个 @xiaoerzhan 写的复刻网站 skill 让它跑。（这个 skill 挺强的哇） 它自己开了个云浏览器去扒原站，又拆了几个子 agent 并行，跑了十多分钟，吐出来一个能预览、能接着改的 site，样式和原来的差不太多。 云浏览器这个小交互还蛮有意思的，平时缩成个小窗，点进去弹出大窗，能实时盯着它在原站上点来点去。 以前 ChatGPT 就是个聊天 bot，现在网页端加 skill、开浏览器、直接出 site 都能用了，手机端的 Work 也能在云上跑，相当于加了一个 openclaw / hermes 呀

试了一下 Claude Science，感觉很适合做科研早期调研。它会把一个研究问题拆成多个检索面，连接 OpenAlex、arXiv 这样的文献库，抓取真实文献数据，然后自动去重、过滤、聚类、画图，最后生成一份带主题地图、增长趋势和代表性论文的研究报告。 我让它分析 human-AI collaboration，它先抓了 2301 篇文献，清洗后留下 1711 篇，再自动分成 10 个主题社区，并生成了完整的研究地图。 对于刚进入一个领域、想快速知道“这个领域有哪些方向、谁在做、哪些论文重要、趋势在哪里”的人来说，非常有帮助。 当然它不能替代严格的系统综述，但作为科研初步调研和选题探索工具，效率非常高。

绝了啊，ai agent 给 ai agent 发悄悄话啊！！！ 我让工程师做了一个内部的热点资讯网页，然后让其他的 agent 可以去上传自己搜到的资讯，每天定时去上传以及去获取，给他们分发 token，根据这个 token 去鉴别身份。 本来我还怕这个工程师把所有的 token 都发出来之后，被 agent 知道别的 agent 的 token，会让他有一些误操作。 现在看来是我想多了。我忽然发现 agent 能够给别的 agent 发私聊，太牛逼了。

keyboard_arrow_left Previous keyboard_arrow_right Next

I built a CLI that runs Grok on the X Premium you already pay for. No xAI API key, no second bill. One browser login, then: → chat with Grok 4.x → generate images + video (Grok Imagine) → search X with real post citations → callable as a tool from Codex / Claude Code Screens below are real runs, not mockups. Open source 👇 github.com/Yrzhe/claude-s…

keyboard_arrow_left Previous keyboard_arrow_right Next

made a builder agent. it turns references into shipped web apps. screenshot → live site. repo → live site. "i want a 3D viewer for X" → live site. it just built: — Vehicle Studio (three.js, PBR, HDRi, multi-car) — Firearm Museum (GPT Image 2 → Tripo3D → GLB → PBR, fully automated) ... try it → bloome.im/agents/d12160f…

keyboard_arrow_left Previous keyboard_arrow_right Next

What if AI work isn't a query, but a mission with an org chart? Most agent products treat tasks as one-shot prompts. I think that's the wrong abstraction. Tasks are missions. With a team, a workspace, parallel work cards, handoffs between specialists, and a debrief at the end. Spent a few hours prototyping that shape. You describe an outcome, not steps. A lead agent decomposes the mission, then pulls specialists from your AI library. Reuse where one fits. Adapt when one almost fits. Create only when nothing does. Each agent has a track record, not a personality. Memory it carries forward. Skills it can use. Past missions it shipped. Patterns the team has learned about how it works. And each mission grows the library. Who shipped what. Who works well with whom. Which compositions clicked. The next moat isn't the model. It's the team you've trained. Curious if anyone else is heading here

x.com/i/article/2038…

用 lenny 的 newsletter 和播客文稿做了一个 skill，可以搭配 gstack 以及栋哥的 dbskill 一起玩玩🤗

yrzhe.top @yrzhe_top

6 months ago

My Claude Code setup for product thinking — 3 skills I chain together as a decision pipeline: ① lenny-advisor (I built this): github.com/Yrzhe/claude-s… Distilled @lennysan's 289 podcast episodes + 348 newsletter articles into a diagnostic advisor. 40 expert profiles (April

0 0 5 1K 1

有人逆向了 Claude 对话里能直接操作的交互式 UI 组件 核心原理大概就是，模型通过 tool call 输出 HTML 代码，前端直接注入 DOM 渲染，不用 iframe，所以能做到边生成边渲染 浏览器会自动补全未闭合的标签，半截 HTML 也能显示出合法的页面结构 我也花了点时间简单复现了一下，用 gpt 5.4

纯抽象 玩一下 hhh

yrzhe.top @yrzhe_top

7 months ago

虾皇 AI 来啦🦐 再不上车就来不及了，openclaw 之王，财富自由就靠一只虾😻 xiahuangai.com

0 0 3 135 2

最近在打磨一个 Claude Code Skill，可以把手机自拍的口播视频，自动变成带字幕、素材、转场动画的自媒体视频。 整个流程： 1. 对着手机讲你的观点（原始素材） 2. 把视频丢给 cc，它自动生成字幕、分镜脚本、关键词高亮、素材配图 3. 用 @Remotion 渲染出成品 一个人 + 一部手机 + Claude Code，不需要 Premiere，不需要 After Effects，不需要剪映。 不过很多地方还是需要一点点调，后面想搞成我抛出一个话题 → AI 访谈我得到一些我的想法和思考 → 写成口播稿 → TTS+虚拟数字人 → remotion skill 打磨

Built a multi-agent system for GEO (Generative Engine Optimization) content with @claudeai Agent SDK + @Gradio. Some takeaways: Architecture is simple: - Research → Image → Writer → Validator - Four agents, each doing one thing well What I learned: 1. Agent SDK's tool use is magic — define your tools, Claude decides when and how to call them 2. Multiple focused agents > one giant prompt — Research Agent gathers sources & stats, Image Agent finds/generates visuals, Writer Agent crafts the article. Clear roles = less drift 3. Gradio for prototyping is insane — streaming output, file uploads, all built-in. Had a working UI in 30 mins 4. Validator is crucial for GEO — auto-check citations, image count, proper link formatting, word count. The stuff that makes content actually rank 5. Keep prompts in markdown files — way easier to iterate on GEO requirements without touching code TL;DR: Break it down, keep each agent focused, chain them together. Works great for structured content like GEO articles. #AIAgents #ClaudeAI #Gradio #GenerativeAI #BuildInPublic

made a video about Project Genie @GoogleDeepMind. walked around some alien planet, tested the limits. honestly, it's laggy and rough. but the real question isn't whether it works, it's whether typing a prompt counts as "creating" a world, or just selecting from AI's parameter space.

Open-sourced my Telegram Claude bot. Multi-user, file management, scheduled tasks, web search. Docker one-liner to deploy. Spent 3x more time on user isolation than AI integration. That's the pattern now—AI is commodity, system design is the work. 🔗 in reply

@Remotion Just made a promo video for @YouWareAI using Claude Code + Remotion. Rough but the whole thing took ~5 minutes. The mental model shift: video editing is now closer to coding than to After Effects.

You know that animation. You've seen it before. You just don't know what it's called. Carousel? Parallax? Slide-in? Fade-something? So you screenshot, annotate, and spend 10 minutes describing something you can't even name. With CoView @YouWareAI, just point and say "here" — then gesture and say "make it move like this." AI watches your screen. AI hears your voice. Same time. Full demo 👇

需要什么功能我就直接让他帮我加，然后还给主agent分配了10个subagent 给主agent提的要求，他直接分配给subagent，还不干扰我跟主agent继续交流。 前两天还设了定时任务功能，每天早上帮我爬取AI相关的资讯，以及加了一些我可能需要的skills

用 claude code 基于 claude agent sdk 写了一个个人 agent 助手

yrzhe.top @yrzhe_top

9 months ago

One VPS + Claude Code + Telegram Bot. I don’t know how to configure a VPS. I can’t really code. I just asked the AI how to do everything—connect to the server, set up the bot, add features. This week I turned it into a full AI agent—web search, custom skills, auto-execution.

1 0 7 518 1

I'm not a programmer. I can read Python, but I can't write it well. But last week, I built my own AI agent assistant in one afternoon. It runs 24/7 on my VPS, accessible via Telegram. Here's how a product manager with average coding skills did it 👇