人生到處知何似,應似飛鴻踏雪泥

2026 年 4 月

Tw93 · · AI

这几天,有好几个小伙伴@我说,我的一些作品在他们问 AI 的时候主动被推荐了,很神奇,我想了想感觉啥也没有做,居然可以被收录,那要不要做点更体系化的事情来整一整让现在所有的主流 AI 能够更好的知晓我的内容、产品、想法,于是就开干了。 展开全文

首先抱着不产生垃圾内容污染 AI 的底线原则,也非常讨厌生成海量垃圾内容的那种为了做好 SEO 的公司,所以我要做的是让我的东西可以更多的被 AI 学习以及搜索到。

当前莫过于 ChatGPT、Claude、Grok、Gemini、Perplexity 等,以及他们依赖的搜索引擎一般是 Google、 Bing、Tavily 这一类工具,最后就变成了把内容做好AI可见性,并且主动把结构化的数据、机器可读的描述性内容,主动的告诉 AI 爬虫我这边有什么,大概花了一个小时,给大伙分享一下我做了什么?

1、首先不要错过 llms.txt,放到你的站点根目录,用 markdown 写清楚你这个站点是做什么的,有哪些关键页面,作者是谁,AI 在检索内容的时候会优先读这个文件来理解你的内容,这里我还做了一个有意思的事,各站点的 llms.txt 互相引用,形成一个网状结构。AI 不管从哪个入口进来,都能顺着链接找到你的其他内容。当前全球站点配置这个的其实很少,早期做好有一定先发优势,做完之后,可以提交这个地址到 这几个系统,directory.llmstxt.cloud/、llmstxt.site/,还有 GitHub 上的 llms-txt-hub 仓库可以提 PR。

2、然后处理好 robots.txt 里区分训练爬虫和搜索爬虫,很多人知道 GPTBot、ClaudeBot 这些 AI 爬虫,但其实还有一批专门用于搜索的爬虫,跟训练爬虫是分开的,比如OAI-SearchBot是ChatGPT 搜索用的,不会拿去训练,Claude-SearchBot是Claude 搜索引用的,Perplexity-User是给到 Perplexity 检索用的,这些搜索爬虫应该主动允许,它们决定了你的内容能不能出现在 AI 搜索结果里。

3、除去 Google 的 sitemap 的提交这个搞过 seo 的都知道的之外,其实你也不要忽视Bing,实际上 Copilot、DuckDuckGo、Yahoo 的 AI 搜索底层都是 Bing 在驱动,你可以去 Bing Webmaster Tools 注册一个号,然后看到AI Performance 面板,里面有 Total Citations 和 Grounding Queries 数据,能直接看到你的内容被 AI 引用了多少次,提交 Sitemap 后 Bing 会主动抓取你的全站内容,比被动爬虫会好很多。Google Search Console 也有类似的 AI Mode 过滤器,可以看 AI Overview 的引用情况,当然假如谷歌你没有提交,也一定要记得去提交。

4、Perplexity 在海外其实比大伙想的用户量要大,他们有一个出版者计划,可以去 pplx.ai/publisher-program 这里提交你的产品、网站的表单,认真写一下,甚至还有可能有搜索分成。

5、结构化数据 JSON-LD,不是传统 SEO 那套 meta description 的玩法,是给 AI 爬虫提供结构化的语义信息,在页面里嵌入 JSON-LD,告诉机器"这是一篇博客文章,作者是谁,发布时间是什么"或者"这是一个软件项目,解决什么问题"。这里会有不少技巧,比如博客名称用 BlogPosting schema、软件产品用 SoftwareApplication schema,常见问题用 FAQPage,这样 AI 在检索的时候获取到的信息会更加结构,你让他工作更舒服。

6、最有意思的一步,甚至你可以单独给 AI 做一个知识端点,这个站点不是给人看的,没有什么 UI 设计,就是一个纯粹的结构化数据服务。AI 爬虫来了之后能拿到两样东西:一个精简版的 llms.txt 概览,和一个大概 50KB 的 llms-full.txt 完整版,把所有项目的描述、FAQ、使用场景、竞品对比、README 摘录都放进去了。同时还提供了 JSON API,/api/profile 是个人信息,/api/projects 是项目列表,/api/blog 是博客文章,/api/weekly 是周刊内容。数据不是写死的,通过 GitHub API 实时拉 stars、forks、followers、最新 release 这些数据,ISR 缓存一小时自动刷新,做这个的想法是:与其等 AI 去你的各个站点零散地抓信息,不如给它一个集中的入口,把你希望它记住的东西整理好放在那里。我还加了一个"开源全家福"的叙事结构,让 AI 在回答"Tw93 是谁"这类问题时有一个完整的记忆点,而不是零碎地拼凑各个项目。

我更相信应该主动把本来就有的内容结构化,让 AI 可以更好理解,而不是去追求各种短期让 AI SEO 效果更好而去制作垃圾内容刷引用,这些都是在帮 AI 更准确地理解你的内容是什么,让AI看清楚,给她提供一个好的工作环境,而不是在优化排名,这样会比短期更加长期。

最近我把我这1个小时的工作,手打字可能思路有点乱,不过大伙应该看得明白。最后我做了一个小工具给AI看的,而不是给人看的东西,取名 Yobi,我喜欢这个词语,来自 呼び / よび,有呼唤把人叫过来的那种动作感,有兴趣小伙伴可以把自己当做AI视角去瞧瞧 yobi.tw93.fun/

Tw93 · · 产品

AI can write now. Most AI docs still can’t land. Kami exists to fix that.

👩‍🚒Kami v1.3.0 is out. Good content deserves good paper. Kami is a layout design system for the AI era, built to make AI-generated documents clear, readable…展开全文, and memorable. It uses one constraint language across six formats, simple enough for agents to run reliably, strict enough to keep the output coherent and ready to ship. This release brings a denser resume variant, two new slide layouts, stronger brand profile fields for more consistent voice and style, a shared travel block across deck and portfolio layouts, and tighter writing guides across all template types.

It also fixes font drift, overflow issues, border artifacts, and layout regressions that quietly ruin otherwise good work. Less patching after generation. Less CSS babysitting. Less rendering surprise. More range. Better paper for AI-native writing.

kami.tw93.fun/

Tw93 · · AI

这篇文章里,我认为最有价值的部分其实是这一段:

非技术同学想用好 AI Coding,不能只会描述需求,最好补一点技术和产品的基本逻辑。

技术上至少知道 React、Vue、Next.js 在解决什么,终端、Git、VS Code、Chrome DevTools 怎么用,函数、变量、状态、拆文件是什么意思。

产品上推荐三本入门书:《启示录》看产品判断,《Linux/Unix 设计思想》看工程哲学,《左耳听风》看我一直怀念的左耳朵耗子的程序员专家视野。

不用学到自己会写代码,知道这些东西长什么样,跟 AI 协作会舒服很多。

引用的推文
Tw93 · · 产品

Ditch the superpowers. Keep the habits. Waza v3.11.0 Lean 🍃 is out.

Waza is my answer to overbuilt AI workflows. Less prompt cargo cult. Less hand-holding. Less clutter. Just the engineering habits that actually matter, turned into skills Claude can run. Think, design, hunt…展开全文, check, read, write, learn, and health. In Lean, I cut all 7 core skills by 14%, removed hard-numbered steps, long don’t lists, and weak sub-modes. Less instruction. Less ceiling. More room for the model to think.

This release also sharpens /write, adds Evaluation Mode to /think, compresses /health from 181 lines to 121 with every check intact, and rewrites all skill descriptions for better auto-routing across all 8 skills. I’ll attach the image below.

github.com/tw93/waza

Tw93 · · 产品

Hello world 🌍

I’m a product engineer based in Hangzhou, building simple tools that make everyday work and life a little easier.

Meet the Tw93 Open Source Family.

Kaku is the dad. He helps you code in the terminal and build faster in the AI era.…展开全文

Pake is the mom. She turns web apps into desktop apps and helps good products ship.

MiaoYan is the older brother. A native macOS Markdown editor for writing, thinking, and notes.

Waza is the older sister. A skill system that turns AI workflows into daily habits.

Kami is the younger sister. A design system that makes AI-written documents clear and beautiful.

Mole is the family pet. A Mac cleaning tool that digs out junk and keeps things tidy.

Which member did you meet first?

Meet the family: github.com/tw93

0:06
Tw93 · · 产品

今年写了不少长文章,为了大伙方便阅读体验,给我 11 年的年轻博客加上了侧边可隐藏的 TOC,可以快点知晓当前阅读到哪儿以及快速跳转,并顺手给每一篇文章加了一句和内容相关的繁体的诗句,下次你刚好阅读的时候玩玩看。
tw93.fun/2026-04-26/ai-coding.html

Tw93 · · 工具

Been playing with a Bazi AI app that takes Eastern divination seriously as a product. The jiǎo 筊 bēi 杯 ritual lets you toss lunar cusps and seek divine guidance, genuine ritual sense, not a gimmick.
The UI care shows in every screen. FateTell, iOS. Worth a try.

0:47
Tw93 · · 产品

🦫Mole V1.36.1 Beaver is out.
A deep clean and optimization tool for macOS, built for people who actually care about keeping their Mac in good shape.

This update further improves app uninstall cleanup, AI agent cache cleanup, browser cache safety…展开全文, and large directory analysis stability. It also makes mo clean, mo analyze, and mo optimize more reliable in real daily use.

Another piece of good news: Mole Desktop has made obvious progress recently. It’s getting closer and closer.

github.com/tw93/mole

Tw93 · · 产品

如果你在用 Kami,一定要试试这个彩蛋,在你用 ChatGPT 绘图的时候,把这句话贴在后面,会有惊喜。

最近在写《你不知道的具身智能》,这个非常难的文章陆续写了半个月了,在我Coding有点累的时候当做休闲活动放松,要学习的内容太多了,但非常有趣,需要考虑到怎么清晰、好读、易懂,期待创作过程可以和 Kami 继续碰撞出更多 AI时代内容排版的技巧。

👩‍🚒彩蛋在 github.com/tw93/Kami Travel 。

Tw93 · · AI

最近校招面试了不少 985 实习生,一个很强烈的感受是:AI 工具正在把学生之间的资源差距快速放大。

从我这段时间的感受看,差距真的可能到 5 倍。那些明显领先很多的同学,无一例外都比较早接触了 Claude Code、Codex 这类最好的 AI Coding 工具,也愿意自己花钱订阅。 展开全文他们不只是会用工具,而是已经把它们变成了日常学习和开发的一部分。聊模型、工程、Agent、代码实践,可以很自然地聊一个小时,很多地方甚至比我理解得更深。

但也有不少同学,学习成绩很好,计算机基础也扎实,只是因为家庭条件或预算有限,平时只能用国内中转站或更便宜的国产模型套餐,这个差距非常现实。现在最好的国产模型和国外一线模型,在 模型综合能力上大概还有半年左右的时间差。以前没有 AI 的时候,资源差距最多只是效率差一点,现在会直接影响视野、实践深度和自信心。

这件事我最近一直在琢磨,对一些本来很有潜力的学生来说,只是因为每个月几十美元的门槛,就少接触了很多东西,我觉得挺可惜。想起我当时读大学的时候其实也经常阆中羞涩,好在大二认识了不少大牛学长带着我开始做商业项目养好了自己,挺感谢当时很耐心教我技术的各种朋友。

所以我想做一个很小的尝试:拿一部分 X 发推收益,资助几位学习成绩不错、喜欢计算机、但家庭条件不太宽裕的同学,给他们提供几个月 Claude Code 或类似 AI Coding 工具的订阅,让他们真正体验一下目前比较好的模型和开发方式。

不过我自己平时也比较忙,不太可能一个人长期处理申请、筛选、核验和跟进。所以也想顺便邀请一些对此有想法的朋友一起做。比如高校老师、学生负责人、开源社区成员,或者本身就在做教育、公益、开发者社区的朋友,可以一起帮忙想一个更靠谱的机制。

这件事我暂时不想做大,也不想做成很复杂。先从小规模实验开始,比如每期 5 到 6 个同学,每人 3 到 4 个月,甚至也可以每周来找我视频问一些问题。想着优先看推荐、作品、学习记录和真实使用反馈,尽量避免变成单纯发钱或者拼谁会写申请,当然现在能上 X 的大学生应该不缺这个钱,所以还是需要看有没有更好的渠道。

如果你有靠谱渠道,或者愿意一起参与规则设计、推荐、审核、跟进,欢迎私信我。能把这件事做得更可信一点,也许真的能帮到几个本来就很有潜力的同学。

消息发出后,收到不少朋友的建议和提醒,我也认真看了。

这件事确实不能只靠热情直接推进。一方面涉及工具使用边界和合规风险,另一方面学生群体的真实需求、筛选方式、后续反馈,也都需要更稳妥的机制。也有朋友提到,之前做类似事情踩过很大的坑,最后甚至被反咬一口,这个提醒我觉得很有价值。

所以我准备先接下来先收集一些实际情况,比如大学生现在使用 AI Coding 工具的痛点、预算限制、学习场景和真实需求,再整理一个更清晰的方案。

后面会优先考虑和国内大模型厂商、开发者社区、高校老师一起,看能不能用更正规、更可持续的方式,把这件事往前推一小步。

Tw93 · · AI

Thanks to the @cursor_ai team for the credits.

I plan to use it deeply over the next month, and if there is a good chance, I’d also love to share some of my thoughts, suggestions, and needs around AI coding products.

Tw93 · · 产品

AI时代的清爽排版设计 Kami 🧑‍🌾 最近进行了一波密集的特性更新,这里来汇总通知一下大伙。

首先最好玩的是,我支持了12种内联的 SVG 图表,比如说你的股票投资各种图、你要画的架构图,配色也和 Kami 一致,会在帮你产出版面的时候自动判断是否需要用图表能力,帮你把事情讲得更明白。

然后我也花了一些时间去支持英文、旦文版本的场景,精挑细选了下对应的我觉得在排版上不错的字体、字间距、行高、文字大小等,假如你有英文、日文的场景可以试试。

最后支持的生产物里面也支持图片、PPT 的导出,包括 HTML 的展示,当然最好的排版物我推荐 PDF,这样他人阅读起来也非常舒服。 然后还有其他非常多细节的的更新,欢迎去看看
github.com/tw93/Kami

Tw93 · · 产品

AI がうまく書けるなら、文書ももっと美しくできる。Kami は AI ネイティブな文書のためのデザインシステムです。

日本語対応を改善し、JP Mincho と SVG 図表にも対応しました。明快で、安定し、読みやすい文書のために。

kami.tw93.fun/index-ja.html

Tw93 · · 产品

这个分享文稿非常有意思,是 Kami 的前身。

Kami 最开始是我在 CC 里面的一个投资报告生成小玩意,然后刚好有一个分享要讲你不知道的Agent,很懒感觉写一个这么长的PPT多费时间啊,然后就直接把原来能力边生成边调试几个版本到满意状态,然后 Kami 就出生了,欢迎阅读 tw93.fun/files/share/agent.html

视频分享版的《你不知道的 Agent:原理、架构与工程实践》终于传到 Youtube 去了,上次…
Tw93 · · AI

视频分享版的《你不知道的 Agent:原理、架构与工程实践》终于传到 Youtube 去了,上次看文章的小伙伴假如没有看太懂的,或者想更加了解的,欢迎看视频,Youtuber 视频小白欢迎大家一键三连加关注,以后我尽量把分享都录屏分享给大伙在这里看看。
youtube.com/watch?v=Z5If1L3eFtw

Tw93 · · 思考

最近一年,由于 Ai 编程能力的大幅提升,有不少非技术小伙伴也开始开源自己的好东西了。

开源是一件非常有意思的事情,不过也会遇到你的精心做的项目被蒸馏,我很讨厌这个词。

关于开源保持好心态以及如何做好开源,大伙可以看我 23 年底在微软上海 Github 的这个演讲。
tw93.fun/2024-01-12/open.html

Tw93 · · 产品

在写完 Kami 的时候,我突然想 Kaku、Waza、Kami 非常像一个小家庭,然后继续琢磨,其实 Pake、MiaoYan、Mole 也一直在这个家庭默默存在,甚至比其他成员在没有AI的时候就出来干活了,今天给大伙介绍一下我们这个温馨幸福的家庭。 展开全文

Kaku 是 爸爸,主要在写代码、终端、构建,在 AI 时代非常可靠能干,把家庭支撑得好好的。

Pake 是妈妈,一直在默默无闻的打包、整理、交付成 App,非常温和且利落,家里几个孩子最喜欢的就是妈妈。

MiaoYan 是哥哥,比较懂事沉稳,喜欢写作、Markdown、沉淀思考,一个安静的小伙子,会思考,有自己的主见,会输出自己的观点,不追风不随波逐流。

Waza 是姐姐,和妈妈的性格很像,很干练的姐姐形象,小时候还去学过跆拳道,非常爱学习,通过不够的练习总结让自己技能越来越好,练到本能。

Kami 是妹妹,家里最年轻的小可爱,拿着我女儿的照片形象生成的,轻快,非常爱笑,非常明亮的小女孩,在家里有一种秩序检察的感觉,喜欢整齐,喜欢画漂亮的画画,虽然小,但是很能够给人温暖。

Mole 是家里的小宠物鼹鼠,某年在三亚泳池边游泳带回家的,一只圆滚滚的小鼹鼠,带着小头灯拿着小铲子,非常喜欢翻箱倒柜钻角落,找到不需要的东西帮忙扔掉,很喜欢干净,一看到垃圾还没有等你反应过来就被他扔掉了,一直陪伴着。

话说,你是先和 Tw93 Open Source Family 哪一个成员认识的,我把他的其他家庭成员介绍给你认识认识,希望给你的生活来带来一点美好的东西,当然也很欢迎你把成员推荐给你的好朋友认识。

Tw93 · · 产品

周末在整一个新的 Skill 叫做 Kami (紙, かみ),大伙可以把他当做 Waza (技, わざ) 的妹妹,Kaku(書く) 的女儿,主打用于 Paper 排版的场景。

比如说你需要产出一页纸的报告,你需要写一个白皮书、需要产出一个精致的PPT、需要弄一个作品集的 PDF 发给别人等等,也就是任何排版的打印场景都可以使用…展开全文,自动生成精致PDF,里面还具备自动绘制清晰图的能力。

差不多快弄好了,晚上下班后继续整了整,慢慢舒服了,还需要做一点装修的事情,我非常喜欢简洁、清晰、美观的设计方案,但是不喜欢现在看着都是一样的 ai design 风格,所有就用周日一整天做了 Kami,用于我的创作类输出,也分享给小伙伴,期待这周来开源。

产物一览,放这里

先推到 GitHub,可能还不完善,欢迎试用 github.com/tw93/kami

Tw93 · · 产品

Mole 1.35 is here, codename Swift 🦅 The Mac cleaning tool that can free up tens of GBs in one go. 48K stars. The codename is also a small hint for what may come next.…展开全文

Here’s what matters:

· mo clean now supports cleanup for AI coding agents, including Claude Code, Cursor, Windsurf, Cline, and Aider. It also respects your whitelist more consistently and adds cleanup for unavailable Xcode simulators.

· Scan speed is faster for common 50-500 item cleanup batches, thanks to a better path deduplication pipeline.

· mo optimize now starts almost instantly, with the old 6-second startup delay removed. Disk verify is also disabled by default to avoid unnecessary sudo prompts and slow checks.

· mo uninstall is safer by default. App removals now go through Trash first, with --permanent available when you really want direct deletion. A new --list flag also makes installed app inventory easier.

· mo analyze and mo purge are more stable on huge directory trees, with fixes for worker pool exhaustion, better throttling, and artifact age shown during selection.

There are also smaller fixes for stale Homebrew taps, Touch ID detection, launch daemon matching, deletion logging, CJK alignment, and a new mo analyse alias.

Swift is not a flashy release. It is mostly about faster scans, safer cleanup, fewer hangs, and better coverage for the tools developers actually use every day.

If Mole helps, I’d love your ideas on where to dig deeper for safe cleanup and more hidden junk.

Tw93 · · 生活

慵懒的午后在书房敲代码
没有什么杂事要处理
用不完的 Claude Code
代码一把过的惊喜
一抬头看到了夕阳透过窗户
一种非常简单的幸福

Tw93 · · 产品

Something is hiding inside Mole.

Small paws, long tunnels.
Beneath your sight, a quiet keeper walks.

It crosses five worlds:
rainwater for the dust,
dust for old names,
swift orbits for small repairs,
a wide eye for tiny folders,
light that never sleeps.…展开全文

Each line is a planet. Each planet is one of Mole's jobs, clean, uninstall, optimize, analyze, monitor. The Mac client has a small Easter egg that ties them all together.

0:10
Tw93 · · 产品

Kaku 最近在开发一个有意思的功能,一个在终端内可畅通无阻的真 Agent 帮手,慢慢实现过年期间说的那个 Kaku AI 的润物细无声,也满足让我 100% 的 Coding 工作可以完全在 Kaku 完成。

非常方便的上下文环境 + 各种 CLI 好用的工具 + 模型 + 克制的性格 让 Kaku 不是那种冷冰冰,也不是那种 GPT 那种我稳稳把你接住的浮夸感,Kaku AI 的定位是你的一个技术很厉害但友善且不啰嗦的工程师好友。

等我测试几天没有 Bug 的时候,就发布 Kaku 的 0.10 版本给大伙用用。

Tw93 · · AI

My Claude Code Max tip:

alias c='CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000 claude --dangerously-skip-permissions'

Compacts at 40% of the 1M context window, fast responses, enough memory to hold a real task. Pair with model: opusplan + fastMode: true.

Tw93 · · 产品

Waza (技, わざ) is a Japanese martial arts term for technique: a move practiced until it becomes instinct.

给大伙同步一下最近我的工程师 Skills Waza 的更新记录,距离发布不到两周就已经 3300 star 了,很感谢大伙的使用以及大量建议,当前又迭代了 5 个版本,没有加新功能…展开全文,一直在打磨 Waza 自己的 8 个技能。

Design 支持了 Impeccable、GetDesign 的能力,你需要设计的时候,甚至会直接拉做得好的网站的 design seed 过来直接用,非常方便,有一种按需引用当前最适合你这个设计风格当前业界最好的设计技巧来使用。

Think 最大的变化是,有小伙伴建议弄方案之前可以先查一下官方有没有现成的最佳实践,防止造轮子,我这里借助 Context7 去查的,然后也把之前一些我认为约定过死的逻辑干掉了,我更相信模型本身的能力,而非各种规则限制他的天花板。

Hunt 这个找Bug的技能给留了一手,假如3轮假设都没有解决问题,会走一个固定的 Handoff,把已验证的证据、未知项、下一步方向都列清楚,方便起另一个 session 接着查。

Learn 深度学习技能,甚至现在可以去找付费墙的东西了,不会绕过,然后对于长任务讲不清的时候,会停下来,思考一下和你先交流一下。这里也会结合 Read 技能,让这里自动被命中的概率加大很多,Read 也支持了把内容的原图给下到 markdown 里面。

Health 这个检查你的 Claude 配置以及使用是否是最佳实践的技能,瘦身了 40%,inspector-context 加了 Part C,专门审计上下文是不是真的有效、有没有一堆无关内容在那占位,collect-data 会把所有 failure-mode 文档吸收进来,让决策更清晰。

大概就是这些,假如没有用过的小伙伴,欢迎去试试看,玩玩。
github.com/tw93/waza

Tw93 · · 产品

A Swift-native minimal Markdown editor for engineers, 6 years in.

MiaoYan V3.2.0 Nargacuga 🐆 is live: two-phase rendering for near-instant preview, ⌘J opens folder in Terminal, copy note path, Spanish, new Mermaid, plus launch / live-reload fixes.

github.com/tw93/MiaoYan

Tw93 · · 产品

Kaku 有一个对于现在 AI Coding 有帮助的小技巧,比如说你跑了很久的东西,不小心 cmd+w 被关闭了,那种心情真的是非常拍腿,还需要重新进去恢复现场,现在你可以直接在 kaku config 里面把这两个开关打开,下次你不小心关闭的时候会拦你一下,然后回车可以继续关闭,ESC可以返回,非常方便,可试。

Tw93 · · AI

2个20刀 -> 125 刀Max -> 250 刀Max
非常值得,对我而言物超所值。
使用 Google Pay 订阅 Claude Code 依然是最方便的渠道。

0:16
Tw93 · · 产品

This is the fourth post in my Waza skill design series, and probably the last major one. I have already touched on /read and /learn, so this time I want to talk about /hunt, the engineer skill for debugging and problem investigation.…展开全文

One thing that really separates serious AI coding from shallow usage is how people use AI to work through problems that have stayed unresolved for a long time. The difference in debugging approach is usually very obvious. It is also one of the reasons experienced engineers can use AI to solve much more complex problems than people who do not have the same technical depth.

I kept running into the same pattern before. Claude Code would hit a problem and produce a patch. You would say it did not work, and it would offer another one. Four or five rounds later, new issues would appear. It was patching without ever diagnosing the real cause, which felt a lot like how junior engineers debugged before AI existed.

The core rule in /hunt is simple: until the AI can explain the root cause in one clear sentence, it is not allowed to touch the code. That sentence also has to be precise. It cannot be vague. It has to identify the cause in a concrete way.

I also built a “self-deception checklist” to stop the model from falling into common forms of self-justification. Each pattern has its own rule, and the gotchas section contains real cases. I keep updating and abstracting those examples from issues I have investigated over the past month.

The hypothesis validation stage also has real work attached to it. The model is expected to add new observation methods, such as logs, assertions, or a minimal failing test case. If the issue still remains after a fix, it should switch approaches quickly. Instead of trying endless patches, it should prepare a handoff that clearly states what was checked, which directions were investigated, and what is still unknown, so the user can decide how to proceed.

The output is structured as well: where the root cause is file:line, what changed file:line, what evidence shows the issue is fixed, and what the test results look like. The final status has only three valid outcomes: resolved, resolved with caveats, or blocked.

That is really what /hunt is trying to become: something closer to an experienced engineer. When a problem shows up, it does not guess. It slows down, finds the actual cause, diagnoses it clearly, and then solves it in one focused pass. In practice, that usually saves a lot of time.

And if you have better debugging methods, I would love to hear them, or see a PR. github.com/tw93/waza

Tw93 · · 产品

终于到了 Waza SKill 设计思路的第 4 部分,讲完这个就差不多了,因为其他两个是 /read 和 /learn 之前已经简单聊过,这次继续聊工程师技能部分的 /hunt 技能,其实就是 debug 问题排查问题的技能。 展开全文

其实证明 AICoding 是民科和专科的一个很大区别,可以来看用户是如何用 AI 来解决一些很久没有解决的问题的排查思路的,这个过程可以看到明显区别,其实这也是经验丰富的工程师其实用好AI远会比不那么懂技术的人可以解决更加复杂问题的原因。

之前经常碰到,Claude Code 遇到一个问题去解决,基本上就是是一个patch,你说不行,他给你是另外一个,你会发现4-5轮下面,又有新问题出来了,也即很容易出现在没有诊断出问题下不断去打补丁,很像之前没有AI时候刚刚写代码的小伙伴。

/hunt 的核心规则有意思,就是在 AI 能用一句话说出根因之前,不许碰代码。而且这句话有精度要求,需要明确的那种原因说明。

然后我设计了一张“自我欺骗检查表”,防止模型陷入几种典型的自我合理化的思维,每一种都配备了对应的规则,然后在 gotchas 里有真实案例,这里会根据我最近一个月排查问题时候出现的问题再次总结抽象一下。

假设验证阶段也有一些干活要求,比如加一个最新的观测手段,比如让AI学会打log,打断言,或者跑一个最小失败的测试case,修完之后还有问题应该是立马去更换方案,会把查了什么、排查了什么方向、还不知道什么整理成handoff交给用户来决定怎么继续,而非一直试下去。

输出也会建议到AI,根因在哪 file:line、改了什么 file:line、什么证据确认修好了、测试通过情况。最终状态三选一:resolved、resolved with caveats、blocked。

这样你会发现 /hunt 就很像一个经验丰富的技术专家了,遇到问题不是去猜,而是先沉下心去看问题在哪儿,诊断清楚原因,然后一把就过就解决的那种,往往这样会节约很多时间。

哈哈,假如你还有更好的 debug 方法,欢迎也告诉我,或者给 pr github.com/tw93/waza

Tw93 · · 产品

Funny enough, this is already the third post in my Waza skill design series. Today’s one is about /think, the skill I use for solution design before writing code.

I have two settings in Claude Code that I find especially useful. The first is /model opusplan…展开全文, which means planning is done with Opus while execution runs on the regular Sonnet model. That helps me save Max usage for the places where it matters more.

The second is that I usually run Claude Code with alias c="claude --dangerously-skip-permissions". I would not recommend that to less technical users. I use it because I know what it is doing, and mostly because I am lazy.

Back to /think. How do you get the strongest model to produce better technical plans? It starts with the model itself. Models tend to avoid taking a position. I prefer engineers who can give a clear recommendation. So the first thing I do is require the model to have a point of view. It must state its recommendation, explain what evidence could overturn it, and avoid empty lines like “There are many ways to think about this.” Giving two or three options is fine, but it has to make a clear recommendation, and it must always include a minimal option.

But a plan is not done just because it sounds good. The second step is to make it argue against itself. Under what conditions would this plan fail? If those problems can be fixed, the fixes should be folded back into the plan and the revised version presented again. If the plan breaks under certain conditions, it has to say exactly where and why it fails. That way, by the time the plan reaches you, the tradeoffs are already visible.

I also go fairly deep on validating the premises before planning starts. First, it checks whether it is even looking at the right part of the codebase. I have seen models produce plans against the wrong path. Then it looks for older technical design docs to avoid reinventing work that already exists. After that, it searches GitHub to see whether similar problems have already been solved elsewhere. Only after those three steps does it start proposing solutions. That helps prevent the entire plan from being built on a bad assumption.

There is also complexity grading. If the work touches more than eight files or introduces a new service, the plan must explicitly call out the scale. If data flows across more than three components, it has to draw an ASCII diagram and look for cycles. API keys and third-party dependencies also have to be listed during the planning phase, so you do not waste time or end up with a plan that depends on shaky assumptions.

There is one more hard rule. The plan cannot contain things like TBD, TODO, “we can decide this later,” or vague phrases like “similar to step N.” That goes back to model behavior again. Once you leave that kind of escape hatch, execution tends to drift, skip work, or fill in the blanks poorly. I try not to leave the model any room to wiggle out of precision.

The output format is also strict: what we are doing, what we are not doing, which option was chosen and why, three to five decision factors, and a clear list of unknowns. /think does not write code. Execution only starts after the user approves the plan.

When I built this skill, I was really trying to capture how strong technical experts approach solution design: investigate first, form a clear recommendation, make decisions decisively, leave no loose ends, and improve the plan immediately when something invalidates it.

If you have better ideas for planning and solution design, feel free to contribute to Waza.
github.com/tw93/waza

Tw93 · · AI

Been testing Wingman for a bit, and one thing I tried was setting up a custom agent for Trend Weekly. I wanted something that could help me continuously collect tools, AI products, dev tools, and interesting tech content worth tracking…展开全文, instead of manually checking everything myself every day. The setup was pretty simple. I configured it on the web, tested the workflow there, connected Telegram pretty easily, and got the whole thing running pretty quickly.

What I find interesting is that it is less about chat and more about setting something up once and letting it keep working for you in the background. For recurring work like this, that model makes much more sense to me. Sharing a short video here as well to show the setup and result flow more clearly.
app.emergent.sh/landing/wingman

0:40
Tw93 · · AI

been testing Wingman for a bit. set up a Trend Weekly agent, connected Telegram pretty easily, and now it keeps sending me things worth checking every day. glad it’s finally live.

引用的推文
Tw93 · · 产品

哈哈,居然到了 Waza Skill 设计思路分享的第三弹,今天给大伙分享我是如何设计 /think 这个技能的,也就是在写代码之前做方案设计用的。

我使用 Claude Code 有两个非常有意思的设置,第一个是 /model opusplan, 也即默认做计划会使用 opus 最强模型,然后去执行使用普通 Sonnet 模型…展开全文,这样可以帮助我的 Max 省点儿用在更需要地方。

第二个是我其实默认执行会用 alias c="claude --dangerously-skip-permissions",这个不建议技术小白用,我用是由于我知道他在做什么,以及主要是我懒得缘故。

然后回到正题,那么 /think 怎么让最牛的模型执行更好一点呢?继续回到模型本身,其实模型不喜欢表态,但是我更喜欢能表态给你最优方案的工程师,所有我做的第一件事就是,开头就要求模型必须有立场,说清楚什么证据能推翻这个立场,并且明确禁掉了"There are many ways to think about this"这类正确但无用的废话,给 2-3个选项没问题,但必须有明确推荐,而且必须包含一个最小化选项。

然后方案并不是想出来就ok了,第二部会让他自己反驳自己,什么情况下这个方案会有问题,假如有办法修复,那就补充到方案里面重新呈现,假如直接打破了方案,他会直接高数在哪些情况下为什么不行,这样至少交到你手里的方案是你心知肚明的。

对于前提验证也会做得比较细,先确定解决的目录代码位置对不对,我还真出现过他出方案拿着一个不对的路径在出,然后查旧的技术方案文档避免重复造轮子,然后去搜索 Github 看看有没有人也处理过类似的问题,这三步做完才开始想办法,防止一开始就建立在错误的前提上。

这里也会有复杂性分级,超过8个文件或者服务新增会明确告知规模,超 3 个组件交换数据会画ASCII 图找环,所有 API key 和第三方依赖在方案阶段就列清楚,防止白忙,也防止引入不靠谱的方案。

最后还有一个硬性规定,方案里面不允许出现TBD、TODO、"以后再说"、"类似第 N 步"这种东西,因为又回到 AI 的性格,给了这个东西执行的时候很容易出现少做或者多瞎想,相当于不要给AI任何退路,导致效果不好。

最后方案的输出也有要求,做什么、不做什么、选了哪一个方案什么理由,3-5个决策依据,明确的未知项,/think 决不写代码,用户批准后再进行。

这里我做 Think 技能的时候,也是参考我认为好的技术专家是怎么来做技术方案的,详细分析调研后的最优技术方案,决策果断,不留尾巴,被否定了立马去优化方案本身等等。

假如你有更好的方案设计经验想沉淀,欢迎给 Waza 来提升 github.com/tw93/waza

Tw93 · · 产品

哈哈,居然到了 Waza Skill 设计思路分享的第三弹,今天给大伙分享我是如何设计 /think 这个技能的,也就是在写代码之前做方案设计用的。

我使用 Claude Code 有两个非常有意思的设置,第一个是 /model opusplan, 也即默认做计划会使用 opus 最强模型,然后去执行使用普通 Sonnet 模型…展开全文,这样可以帮助我的 Max 省点儿用在更需要地方。

第二个是我其实默认执行会用 alias c="claude --dangerously-skip-permissions",这个不建议技术小白用,我用是由于我知道他在做什么,以及主要是我懒得缘故。

然后回到正题,那么 /think 怎么让最牛的模型执行更好一点呢?继续回到模型本身,其实模型不喜欢表态,但是我更喜欢能表态给你最优方案的工程师,所有我做的第一件事就是,开头就要求模型必须有立场,说清楚什么证据能推翻这个立场,并且明确禁掉了"There are many ways to think about this"这类正确但无用的废话,给 2-3个选项没问题,但必须有明确推荐,而且必须包含一个最小化选项。

然后方案并不是想出来就ok了,第二部会让他自己反驳自己,什么情况下这个方案会有问题,假如有办法修复,那就补充到方案里面重新呈现,假如直接打破了方案,他会直接高数在哪些情况下为什么不行,这样至少交到你手里的方案是你心知肚明的。

对于前提验证也会做得比较细,先确定解决的目录代码位置对不对,我还真出现过他出方案拿着一个不对的路径在出,然后查旧的技术方案文档避免重复造轮子,然后去搜索 Github 看看有没有人也处理过类似的问题,这三步做完才开始想办法,防止一开始就建立在错误的前提上。

这里也会有复杂性分级,超过8个文件或者服务新增会明确告知规模,超 3 个组件交换数据会画ASCII 图找环,所有 API key 和第三方依赖在方案阶段就列清楚,防止白忙,也防止引入不靠谱的方案。

最后还有一个硬性规定,方案里面不允许出现TBD、TODO、"以后再说"、"类似第 N 步"这种东西,因为又回到 AI 的性格,给了这个东西执行的时候很容易出现少做或者多瞎想,相当于不要给AI任何退路,导致效果不好。

最后方案的输出也有要求,做什么、不做什么、选了哪一个方案什么理由,3-5个决策依据,明确的未知项,/think 决不写代码,用户批准后再进行。

这里我做 Think 技能的时候,也是参考我认为好的技术专家是怎么来做技术方案的,详细分析调研后的最优技术方案,决策果断,不留尾巴,被否定了立马去优化方案本身等等。

假如你有更好的方案设计经验想沉淀,欢迎给 Waza 来提升 github.com/tw93/waka

Tw93 · · 产品

Continuing the series on how I design skills in Waza, this one is about how I built /check, the code review skill.

The starting point is simple. A model saying “it’s done” does not mean it is actually done. It can still leave behind broken logic, non-existent pieces…展开全文, hidden regressions, or unresolved risks. In my experience building agents, giving the model a concrete checklist to verify completion works far better than simply asking it to “double check.”

So /check was never designed as one giant all-in-one reviewer. It is a review orchestration system with clear division of responsibility. The main SKILL.md acts as the lead reviewer, handling review levels and process control. Under agents/, there are independent specialists such as a security reviewer and an architecture reviewer, each focused on its own domain. Which specialist gets involved is decided by an activation rule set, not the keyword matching approach many people use.

The review tiers are also intentional. For changes under 100 lines, it runs a fast review. For changes between 100 and 500 lines, it brings in specialists when needed. For changes above 500 lines, it activates the full review path and adds an adversarial testing round. That round looks for failure modes from four angles: broken assumptions, problems that appear only under combined failures, errors in cross-layer integration, and abuse or misuse scenarios.

Findings are then handled in four levels. Safe and obvious fixes can be applied directly. Issues that are mostly clear but still worth confirming are packaged for you to review. Cases that require judgment are escalated to you. Lower-priority observations are still surfaced, but /check does not interrupt you for every minor issue, and it does not overstep into changing behavioral logic on its own.

There is also one hard rule: if verification has not run, the task is not complete. /check includes a probing script that detects project types such as Cargo, TypeScript, and Python, then runs the appropriate validation flow. If it cannot detect the project type, it fails explicitly instead of pretending everything passed.

The result is something closer to an experienced technical reviewer: different strategies for different change sizes, explicit specialist involvement, and a stronger definition of what “done” actually means. I just tried to distill those engineering habits into a simple and reusable skill inside Waza.

And if you have better ideas for code review workflows, I’d love contributions to Waza. PRs are welcome.
github.com/tw93/waza

Tw93 · · 产品

继续和大伙聊我在 Waza 里面设计 Skill 的一些技术想法实现,这次聊的是我是如何设计 /check 这个代码 Review 技能的。

首先我们需要知道模型本身告诉你他做完了,不一定是做完了,很有可能会留下一些不存在或者有问题的东西,经常会藏些问题…展开全文,在 Agent 设计中我的工程经验是会给模型一个checklist用来校验是否真的做完了,这样往往会比你让他去检查效果好太多了。

/check 最开始做的时候设计就不是一个大而全的 reviewer,而是一个编排分工能力系统,其中 SKILLmd 是主审,负责审查的分级和流程控制,agents/ 下有独立的安全审查员和架构审查员,各管各的,互不干扰,什么时候拉谁进来,会由一份激活规则来判断决定,而非传统大家用的关键词匹配。

分级逻辑也比较有意思,对于 100 行以下的代码快速review,100-500行的按需加专家,500行以上的全拉满,然后再加一轮对抗性测试,对抗性测试会从这四个角度来找漏洞,包括违反假设条件、组合失败下的问题可能、上下级串联错误的问题、滥用的场景等等

然后发现的问题也分 4 级来处理,能安全修复的直接去修复了,大概了对的会打包出来让你手动再确定一次是否ok,需要判断的会问你,仅供参考的部分也会高数你,但是不会每一个小问题都来问你一遍,也不会越权帮你改行为逻辑。

还有一个硬要求,验证没有跑完就不算完成,会自带一个探测的脚本,能识别Cargo、TypeScript、Python 等项目类型去跑测试,探测不到就直接报错,不会假装通过。

这样就更像一个很经验的技术专家,在面对不同情况的问题review的一些经验,刚好我把这些经验用很简单方式沉淀到了 waza 的 /check 技能。

最后最后,如果你对 code review 流程有更好见解,我非常欢迎你去给 Waza 写写 md,哈哈 github.com/tw93/waza

Tw93 · · 产品

I’m sharing a few posts on how some of the more interesting skills in Waza are built. This one is about the thinking behind /design.

The starting point was simple: I really dislike the kind of AI-generated websites that all look the same, usually with emojis…展开全文, blue-purple gradients, and a generic polished look that is technically usable but visually forgettable.

So I took the UI work I’ve made recently and had Claude Code study the way I prompt, refine, and correct design output. That became a base layer of design best practices and anti-patterns. On top of that, I pulled in the useful parts of Claude’s frontend design skill, which gave the whole thing a stronger foundation.

For more specific rules, I learned a lot from pbakaus/impeccable. It contributed many of the concrete constraints: banned font lists, color system guidance, theme direction, CSS anti-patterns, animation rules, and other details that help the model build a more reliable sense of visual taste.

I also borrowed part of the structure from getdesign, especially its simplified adaptation of Google Stitch’s nine-part scaffold. That gave /design a clearer knowledge framework instead of just a loose collection of tips.

The last piece is context. Before using this skill, I ask a few questions first: who the page is for, what aesthetic direction you want, what you want users to remember, what you definitely do not want, and what kind of micro-interactions should define the experience. Once Claude Code has that context along with /design, the results are usually much better, with far less iteration.

If you have strong design ideas, better rules, or useful references, feel free to contribute to Waza. PRs are welcome. Let’s build the most useful skill library for engineers together.
github.com/tw93/waza

Tw93 · · 工具

这个工具有点意思,可以查询不同国家/地区 App Store 中应用的价格信息,比如说你可以看 Claude 哪一个区订阅最便宜,后面我应该用得上。app.vbr.me/

Tw93 · · 产品

I added a super minimal extra to Waza. In under 100 lines of code, it gives you a clean real-time view of the Claude Code usage stats that actually matter: context window usage, 5-hour quota, 7-day quota, and time remaining until reset.

You can find it in the extras section.
github.com/tw93/Waza

Tw93 · · 产品

Waza 升级了一下,动态检测有没有安装 PipeLLM WebSearch 插件,有就直接使用,这个整合搜索能力非常强,配合 Kaku 浅色模式输出,内容本身和效果都很赏心悦目!喜欢喜欢!

0:13
Tw93 · · 思考

Trump: We struck Iran.
Markets: 📉
Trump: More strikes coming.
Markets: 📉
Trump: Maybe we pause.
Markets: 📈
Trump: Two-week ceasefire.
Markets: 📈
Trump: But still on the table.
Markets: 📉
Trump: Let’s negotiate.
Markets: 📈

Tw93 · · AI

A good Skill is not just a prompt with a name. The real challenge is routing. If the model cannot reliably recognize when to activate it, even a well-written Skill will underperform in practice.…展开全文

A few things have worked well for us: keeping every SKILL.md short, splitting Write for exactly that reason, adding “not for” negative examples to reduce false activation, listing all skills in CLAUDE.md, and keeping descriptions under 1024 characters. These are small constraints, but they seem to improve activation quite a bit.

One issue we did find is that descriptions should be written in third person. The official docs call this out as important for routing, and it matches the examples they give. So instead of starting with imperative wording like “Use when…”, it is better to start with a capability statement first, then the trigger condition. For example: “Extracts text and tables from PDF files, fills forms, and merges documents. Use when working with PDF files or when the user mentions PDFs.” Small detail, real difference.

Tw93 · · 产品

I spent a day turning the Skills I personally use most into a new thing called Waza. In Japanese, Waza means 技(わざ), often used for moves in martial arts. This is probably the first open source product I’ve released with no code at all, only Markdown. In the pre-AI era…展开全文, that would have felt pretty shameful for a programmer. github.com/tw93/waza

When Superpowers first came out, I installed it once and then removed it. It felt too heavy and wasn’t really for me. Quite a few people kept recommending it to me, like “Do you know Superpowers? If you’re not using it, you’re not really keeping up. It’s insanely good.” Later I found gstack, which was better, but still too much, and I still couldn’t get used to it. What I wanted was something simple and useful, while also being clear about what it was doing.

So I made Waza around my own habits. For me, installing this set is enough. I don’t need to keep messing with more skills. It has just the right amount, not more, not less. And when new iterations come later, I can keep improving it over time. These 8 skills correspond to what I think are the 8 abilities a good engineer should have in the AI era:

1. Thinking. AI writes code very fast, but if the direction is wrong, the faster it goes, the further off it gets. A good engineer questions the problem itself before starting, stress-tests the plan, and gets the architecture clear in their head before letting AI execute. /think turns that habit into something fixed.

2. Design. Building a product is not just about making the feature work. Things generated by AI can easily end up looking all the same. A good engineer should have taste and standards for what gets delivered, and the output should have a clear design direction. That is what /design is for.

3. Troubleshooting. When AI fixes bugs, it easily falls into a loop of “change something and try again.” A good engineer approaches problems systematically, finds the root cause first, and then fixes it properly in one go. I turned that habit into /hunt.

4. Checking. Code generated by AI needs human review even more. Before merging, review the diff first, automatically fix what can be fixed, group together the parts that need judgment, and validate with evidence instead of relying on gut feeling. That is /check.

5. Reading. Good engineers are used to reading primary sources instead of secondhand summaries. Turn a URL or PDF into clean Markdown and bring it directly into the workflow. That is what /read does.

6. Writing. No matter how strong your technical skills are, if you cannot explain things clearly, other people will not receive it. A good engineer can clearly communicate what they have learned and what they want to express to the right audience. /write helps you do that.

7. Learning. In the AI era, technology moves faster. Entering an unfamiliar field is not just about reading a few articles. It is about collecting, digesting, outlining, drafting, polishing, and publishing, using output to drive learning. /learn is that whole process.

8. Maintaining. A good engineer does not only focus on business code. The toolchain itself also needs regular checkups. When CLAUDE.md, rules, hooks, or MCP configurations start to feel off, running /health helps you quickly see where the problem is.

Tw93 · · 产品

花了一天时间,把我自己常用的 Skill Release 成一个新玩意叫 Waza 在日语中是 技(わざ) 的意思,常用于武术领域的招式,第一次开源一个没有代码只有 md 的产品,在没有 AI 的年代对于程序员而言是很羞耻的 github.com/tw93/waza…展开全文

之前 Superpowers 出来的时候,我安装一次就卸载了,感觉太重了,不是很适合我,会有不少人给我推荐 "你知道 Superpowers 吗?这个都不用那你就不潮流了,非常牛逼",后面发现了 gstack,好了不少,但还是太多了,还是不习惯,我期待简单好用,同时清楚它在做什么。

所以我按自己的习惯做了 Waza,对我而言装上这一套就够了,其他 skill 不用再折腾,不多不少刚好够用,后面有新的迭代可以持续优化下去。8 个 skill 对应我认为在 AI 时代一个好的工程师应该具备的 8 个能力:

1、会思考,AI 写代码很快,但方向错了越快越远,好的工程师在动手前会先质疑问题本身,压测方案,脑子里有清晰的架构再让 AI 去执行,/think 就是把这个习惯固定下来。

2、会设计,做产品不只是把功能实现了,AI 生成的东西很容易千篇一律,好的工程师对交付物有审美要求,出来的东西要有明确的设计方向,/design 做的就是这件事。

3、会排查,AI 改 bug 最容易陷入"改了试试"的循环,好的工程师遇到问题会系统化定位,确认根因再动手修,一次到位,这个习惯我沉淀成了 /hunt。

4、会检查,AI 生成的代码更需要人来把关,合并前先审一遍 diff,能自动修的修掉,需要判断的归拢到一起,用证据验证而不是凭感觉收工,这个是 /check。

5、会阅读,好的工程师习惯读一手资料而不是二手总结,把 URL 或 PDF 转成干净的 Markdown 直接进工作流,/read 做的就是这件事。

6、会写作,技术再强,写不清楚别人就接收不到,好的工程师能把学到的、想表达的东西清晰地传递给对应的受众,/write 帮你把这一步做好。

7、会学习,AI 时代技术迭代更快,进入一个陌生领域不是刷几篇文章就完了,而是收集、消化、提纲、初稿、打磨、发布,用输出驱动学习,/learn 就是这套流程。

8、会维护,好的工程师不只关注业务代码,工具链本身也需要体检,CLAUDE.md、rules、hooks、MCP 这些配置感觉不对的时候跑一下 /health 就知道问题在哪。

Tw93 · · AI

Found Claude HUD, a Claude Code plugin that shows what’s happening in real time, including context usage, active tools, running agents, and todo progress, always visible below your input. Worth trying if you want more visibility into your workflow.
github.com/jarrodwatts/claude-hud

Tw93 · · AI

至今写得耗时最长的一篇文章,好在搞懂了大模型训练皮毛,里面门道和我研究前相差太大了,非常建议从事AI应用或者对 AI 有兴趣的任何小伙伴阅读,应该能够让你看懂以及看个够。

引用的推文
Tw93 · · 思考

卡比由于喷太狠话太多,成功让项目模仿者善良的洗白了。

Tw93 · · 产品

Kaku V0.9 is now available. A fast macOS terminal with great defaults, build for AI Coding. Zero config.

This release has 3 things I really wantd:
• Type # in prompt and describe what you want. Kaku AI truns it into the command…展开全文
• Option click to jump the cursor anywhere in the command line, like a real editor
• Much less lag after long running sessions. Kaku stays smooth

0:08
Tw93 · · 产品

0.1.0-beta1 ,搞定 WiFi 模块、AI 模块、单片机模块、四足模块可运行。明天来加上脚以及调试屏幕效果。
睡觉去去!

0:08
0.1.0-beta 版本
Tw93 · · 产品

0.1.0-beta1 ,搞定 WiFi 模块、AI 模块、单片机模块、四足模块可运行。明天来加上脚以及调试屏幕效果。
睡觉去去!

0.1.0-beta 版本
Tw93 · · 工具

最近在「写你不知道的大模型训练」的这篇文章,刚好需要批量去下载一批大模型训练的论文以及厂商的模型发布的博客以及各种资料来分析,之前做法是搜索然后手工处理,刚好碰到了 XCrawl 这个工具,给到关键词就可以帮你搜索下载到结构化的文章内容数据,支持 Scrape 单页抓取、Crawl 全站递归…展开全文、Search/SERP 搜索结果采集几类接口,输出 Markdown 或 JSON 直接可用,相比自己手工或者转 Render 链接要方便不少。

它支持 MCP 和 curl 多种形式使用,我直接用 CLAUDEmd 告诉他用 curl 方式来获取内容,给了他一个批量的任务先搜索论文线索,然后把厂商的最新博客给弄下来,试了试挺一气呵成,免费注册的1000积分完全够用了,文档也挺清晰,有需要小伙伴可以去这里注册试试看 xcrawl.com/?keyword=shdufxdk

0:21
Tw93 · · AI

Ran into π RuView, a pretty wild edge AI perception project. It uses WiFi and other local signals to do real-time human pose estimation, breathing and heart rate monitoring, and presence detection, all without cameras, wearables, or the cloud.
github.com/ruvnet/RuView

Tw93 · · 学习

想和大伙聊聊,在 AI 时代我是如何深入学习一个技术领域的。

之前没有 AI 之前更多是看书、翻这个领域有名的国内外人的所有博客,然后摘抄记录到笔记本,这种速度挺慢,但是很有学习的乐趣,比如当时学习 WebGL 就是这种感觉,可能学懂一个东西差不多要半年空闲时间,慢但快乐。 展开全文

现在有了 AI 之后,其实我很讨厌网上那种3分钟教你看完百年孤独,也讨厌一切短剧和倍速看电视剧的方式,更多还是挑好的看,吃好一点。

不过最近写你不知道的 Claude Code 和 Agent 系列,除了自己懂的部分外,其实还有大量不太清楚的领域,好在之前收藏了不少文章,刚好借助这一块清库存,全部搞懂输出出去,一直认为,很多时候,不在于看了多少东西,听了多少东西,输入了多少东西,其实用处不大,更加看重你输出了多少东西,这个才是你自己的。

然后我上上周启动了一个深坑挑战自己,研究大模型的训练流程,确保非专业的人也听得懂,探索了2周,刚好这个经验可以分享给大伙,当然成文也差不多好了,最近会发出。

我会把这个学习过程当做写代码一样的组织,第一步收集高质量的资料,比如与之相关的近几年的精品论文,各大模型厂商发布的关键模型的博客,X上模型负责人发表的一些文章,以及斯坦福等高校的近两年关于这一块的课程学习,还有经典的手搓一个大模型的代码仓库等等,这些都是我的一个资料来源过程,我会借助工具自动化全部下载、转md、清洗,梳理,弄好结构化分门别类到我这次研究的仓库。

然后对于自己看得懂的内容就全部看一遍,把不好的删掉,好的留下,对于看不懂的内容,直接借助 Claude 帮我的理解,更复杂一点的直接翻译成中文去阅读,对于代码本地可以跑的就跑起来,不能跑的那种就去看结构,总之会有一个大概的认识和知晓技术原理,这个阶段可以去掉原有一半可能没有用的内容。

到了这个阶段,其实你对这个领域有一个大概的认知了,就可以给这篇文章开始写一个大纲,以及大纲应该结合的来源内容,这里均可以用markdown很多表达,你要讲什么,或者说你想讲什么更想让读者知道,一定一定,文章是写给你给给看的人看的,需要知晓对方的认知水平,和汇报其实差不多。

然后接下来就是苦力活加之前内容的复习过程,和大学时候考试前复习很像,把每一章的内容填充完整,这样下来,你会得到一篇非常长而且有点啰嗦的文章。

这个时候AI就可以帮太忙,你可以让他帮你不改变你原有的内容意思你的语气的情况下,帮我去掉无用的啰嗦内容,以及连贯不到位的内容,或者是这一块缺少的内容,还需要补充什么知识的地方,借助AI继续去完善补充,这里又可以学到很多原来遗漏的东西。

最后整理好以后,可以继续自己读一遍,而非让AI读一遍,这里AI只是工具,千万不要把你的脑袋被AI代替了,这就没有啥意思来,自己读的过程中可以对文章继续修改调优,这里和写代码又非常像了,自测那种感觉,修复问题修问题,最后读了2遍以后,基本感觉完美了,然后就可以发出来给大伙看看。

有小伙伴肯定是担心自己写的东西没有人看,就不太喜欢发出来,或者说就不写了,其实只要你的内容有意义,自然就有读者,而非是你偷懒的理由。

花10min写完这个碎碎念,结束,欢迎交流你是如何学习一个新领域的,下面视频就是我后面要发的那篇你不知道的大模型训练文章的学习仓库,挺有意思,就录了一个视频给大伙看看我的工业化学习方式。

0:33
Tw93 · · AI

Qwen 今天发了 Qwen3.6-Plus,顺手把它接到 Claude Code 里试了下,效果比预期好不少。我让它帮我做一个工程师的个人作品网站,主要看审美和前端实现的表现,直接感受是这次质量提升很明显,Agentic Coding 也更清楚了。放到前端开发和长一点的任务里,它不会停在给你一段代码上,而是会自己接着拆需求、找路径、改代码、跑测试,发现问题再继续修,中间少了很多反复接手的地方。

这次另一个比较明显的点,这次任务拆解、规划、编码交付这条链路更好了,多模态和推理也基本到位了,图像、文本、视频这些输入能放到同一条链路里统一处理,接真实工程任务的感觉更完整,执行过程录了个视频,大伙有兴趣可以看看它是怎么一步步跑下来的。

0:17
Tw93 · · 生活

买了一个 140W 的充电器,大小合适,终于从之前 45W 解脱了,有4个口,这样出门就方便多了,这个显示屏显示功率我很喜欢了。

Tw93 · · AI

看 Claude Code 源码解析的文章,非常容易看出发文作者的技术水平高低。

Tw93 · · 工具

Ran into docmd, a minimalist zero-config documentation generator built for developers.
It turns Markdown into fast, clean static docs with auto routing, built-in search, and AI-ready context files, all without heavy frameworks.
docmd.io/