人生到處知何似,應似飛鴻踏雪泥
Tw93 · · 产品

Pake V3.18 is out. It's an open-source tool that turns any webpage into a lightweight desktop app with one command, now at 62.9K stars.

What's new:
1. The docs have their own site now, with a step-by-step getting started tutorial and pages for shortcuts, scripts and AI agents.…展开全文
2. Dragging a file onto a page to upload it now always works.
3. Downloads keep their real names, so Annual Report.pdf no longer gets saved as Annual%20Report.pdf.
4. On Windows, downloads show the native WebView2 panel with progress, and a failed build no longer blocks the next one.

Thanks to the contributors whose pull requests went into this release.

github.com/tw93/Pake

Tw93 · · 生活

🤣 笑死居然把这款漱口水在脆上面带火了
已经变成生活博主了吗?

0:29
Tw93 · · 生活

最近上映的这个埃及的电影《生日快乐》很好看,推荐有空的小伙伴一看。

Tw93 · · 学习

People often ask me how to use AI to design products that look good and don't feel like AI slop. It's not an easy question, since I'm not a professional designer, so I can only share it from a product engineer's point of view:…展开全文 how to make products that aren't ugly and are easy to use.

First, one premise. A good-looking design can't save a product that doesn't solve a problem. Design is the icing on the cake, not what decides whether a product succeeds. If I had to rank priorities, it would be: solves a problem > easy to use > looks good. It's like food: first don't go hungry, then get enough to eat, then eat well.

I won't say much about solving the problem, since that depends on the product itself. It solves a concrete need: someone needs it, there's a real scenario for it, and your product happens to offer a way for a person or an agent to get it done. Some problems are just bigger than others.

So how do you make a product that's easy to use and not ugly? Ha, you can walk through the list below yourself, or hand the whole thing to your agent and have it review your product's codebase.

1. Pretend you're an ordinary user who knows nothing about tech. Can they get through the main flow and come out with what they need? Find every place where they get stuck, where something isn't explained, or where one step doesn't lead naturally into the next, and fix all of it, so people can use the product without a manual. Only where that's truly impossible, add a short hint.

2. Every piece of copy should read the way native speakers of that language actually talk. Choose your words carefully instead of translating flatly. Look at how Apple, Windows and well-known websites word the same thing, and aim for copy that feels simple, clear and natural to a local.

3. Where the product itself already makes something clear, don't add text to explain it. Don't over-explain. Check every line of helper text for whether it says what it means without padding, and if the interaction or the visuals already say it, don't say it again in words.

4. When you show a number, make sure people know what it refers to, and make it easy to read for what it means. Money, metrics, distance and other units each have their own best practices. Accuracy matters a lot, so don't skip validating the data to save effort.

5. Don't keep asking users to confirm something they already chose. Ask a second time only when an action can't be undone, so the screen isn't full of confirmation dialogs. Don't splash big red areas over dangerous actions either. Think about how pleasant the product feels to use. Users aren't stupid, so don't worry too much about them making mistakes.

6. Visually, keep one accent color per page. Try to avoid orange, terracotta, and the bright corporate blue of old admin dashboards. Make every gray a warm gray, never a cold blue-gray. If you can't find a good color, look through traditional Chinese colors.

7. Use as few font sizes as possible, at most four on a page. Think about the reading rhythm of the whole page, its clarity and text contrast, and make sure it meets accessibility standards. Line height, letter spacing and paragraph spacing all need repeated tuning, while the whole system stays consistent.

8. Use italics, monospace and all caps sparingly, and only with a clear reason. No italics for helper text, monospace only for code, paths and metrics, and no all-caps text in buttons or labels. Go easy on joining things with a middle dot (·). Use one verb for one action across the whole product, and avoid verbs that push people to act, like the marketing-style "Buy now".

9. Keep buttons, spacing, corner radius, dividers, dialogs, animation and shadows consistent and restrained, so they don't get in the user's way. Nested rounded corners should be concentric: the inner radius equals the outer radius minus the padding between them, and it looks better. Pages are flat by default: no stacked shadows, no glow, no frosted glass. Shadows are only for dialogs that really float above the page, and no cards inside cards.

10. In a multilingual product, lay things out for the longest translation first, and put meaning before looks rather than changing what the text says to make it fit. You can rewrite the copy, but don't change the font size or truncate it. Check that everything displays well at different sizes on the devices your users actually use, test them all, and fix everything that looks off.

11. For a design you keep tweaking without getting it right, or one that still looks off after many rounds, ask your agent for the three options it thinks are best and pick one, then tell it to keep refining until you're happy. I like having AI put before-and-after images into one PDF, which makes comparing them much easier for me.

That's about it. One more thing matters a lot, though: your own patience and willingness to keep iterating. I'm a bit of a perfectionist, and I often go a dozen rounds with AI over a single line of copy, a color or a design until I'm happy with it. AI is smart, but you often need to ask more of it, and put in more time and patience to get a result you like. Or put another way, spend more time and look at more good design, and you can build products that are easy to use and not ugly.

Tw93 · · 生活

假如你在杭州想尝试一下素食
一定要去试试庆春里,很好吃
特别是黑松露炒饭还有藜麦椒麻豆腐

Tw93 · · 学习

经常有小伙伴问我,怎么用 AI 做出好看的产品设计,不那么 AI Slop,其实不是很好答,因为自己也不是专业的,只能站在产品工程师视角来给大伙聊一聊,怎么做出不丑且易用的产品。

首先要知道一个前提,很多时候好看的设计也救不了不解决问题的产品,设计在这里只是锦上添花的作用…展开全文,而非决定产品的成败,假如要我选择产品优先级,应该是能解决问题 > 易于使用 > 好看美观,好比先解决不饿着、再解决温饱、最好解决吃好一样。

能解决问题这里不多说,更多取决于产品本身,它是去解决一个具体的需求,有人需要,有场景需要,刚好你这个产品可以提供一个解决方案让人或者 Agent 使用完成目标就好了,只不过解决问题有大有小。

那么如何做易用且不丑的产品呢?哈哈,可以照着走一遍,也可以把下面的这一大段给你的 Agent 让他去帮你 Review 一下你的产品代码仓库。

1. 把自己当做一个普通不懂技术的小白,他拿到我们产品,能否把主链路走完产出他所需要的东西,哪些地方有卡点,哪些地方没有讲明白,哪些地方的衔接不通顺,全部去修复,让用户不用说明书也可以完成产品的使用,假如实在不能,可以适当补充提示说明。

2. 所有文案需要符合对应语言当地人的使用习惯,需要仔细斟酌词藻,不要干巴巴的翻译,可以从 Apple、Windows 系统以及知名网站去找同类词的最好翻译和解释,文案做到尽可能让本地人感觉非常简单、易懂、清晰。

3. 产品上能用产品本身表达清楚的地方就不要再加文案说明,不要啰嗦的解释一个东西,仔细思考每一个说明文案是否达意且不啰嗦,交互视觉可以说清楚的地方就不要用文案再表达一遍。

4. 在数字的表达上需要让用户明白这个数据指的内容,以及考虑到数据表达意思后的易读,比如钱、指标、距离各种单位需要考虑到本身的最佳实践,数据的准确性非常重要,不要为了偷懒少了数据的校验。

5. 对于用户确定的操作不需要反复找用户确定,只有在会产生不可逆影响的时候二次找用户确定是否需要执行,防止满屏都是各种弹框确定,对于危险的操作不要出现大片的红,需要考虑用户操作一个产品的愉悦度,用户不是傻子不要过度担心会犯错。

6. 视觉上,一个页面只留一个强调色,尽量不要使用橙色、陶土色和后台常见的那种老企业亮蓝色,灰色全部用偏暖的灰,不使用冷冰冰的蓝灰,找不到合适的颜色可以去中国传统颜色里面去找到合适的。

7. 字号类型尽可能少一点,一个页面最多控制4种不同的字号,需要考虑到页面整体的阅读节奏感,清晰感以及文字的对比度,确保符合无障碍,包括考虑到行间间、字间距、段落间距需要反复调整,同时保持整体的规范一致。

8. 少用斜体、等宽字体、全大写,要用的场景必须要有详细的理由说明,说明文案不要使用斜体,等宽字体只用于代码、路径、指标,按钮和标签里面的字不要大写,少用中间点 · 来连接内容。同一个动作在整个产品里面只用一个动词,此外动词的考虑尽量不要催促用户去执行,比如少用立即购买这种营销词。

9. 页面的按钮、间距、圆角、分割线、弹窗、动画、阴影这类设计规范保持一致,同时保持克制,不打扰用户,嵌套的圆角要同心,里面的圆角等于外面的圆角减去间距会更好看。页面默认是平的,不叠阴影、不发光、不加毛玻璃,阴影只给真正浮起来的弹窗,不要卡片套卡片。

10. 在多语言的环境下,优先考虑最长文案的排布,优先考虑达意而非为了好看而改变原意,可以改文案,但不要变字号和截断。需要特别考虑在不同尺寸下的展示效果是否都良好,考虑到用户常用的设备,全部去测试,遇到不好的地方全部修复。

11. 对于反复调整不好的设计,或者做了好多轮都感觉不好看的设计,让 Agent 帮你出3个他认为最好的方案来让你挑选,然后你选择一个更好的,告诉他去优化,直到改到你满意,我比较喜欢让 AI 生成对比前后的图片放到一个pdf里面,这样我的阅读对比体验会更好。

大概就是上面这些,不过还有一个非常重要的就是你自己的耐心和反复调整,我比较强迫症,经常会因为了一个文案、一个颜色、一个设计和 AI 讨论十几轮,直到自己满意了才舒服,虽然 AI 很聪明,但是很多时候你对他的要求要更多,以及你需要付出更多的时间和耐心,才可以做出满意的效果,或者说你多花时间,多去看好的设计,其实就可以做出易用不丑的产品了。

Tw93 · · 生活

疯狂推荐用这个来洗你的茶杯,咖啡杯,或者任何洗不干净的杯子碗

一点点下去,有魔法一样立马光亮如新!

Tw93 · · 思考

Yesterday a reader sent me a message. They study computer science, and they feel AI can now do all of their homework and papers for them. They asked whether this field still has a future, and what advice I'd give for the years ahead.…展开全文

It made me think: if I could go back to my final year of college, what would I do to make the road ahead a little smoother, with fewer pitfalls? This is just my own limited view, and I'd love to hear yours.

1. I'd still choose computer science, then pick the sub-field I find most interesting, and use AI to amplify my work there as much as I can, so that right after graduating it already feels like leading a team.

2. Move to a city with a strong tech scene and plenty of tech companies, like San Francisco, New York, London or Seattle. If your degree and skills are strong enough, find a good AI company in the few blocks of San Francisco between SoMa and Mission Bay. Being that close to the AI labs will sharpen your sense of where large models are heading far more than anywhere else.

3. At work, get things done fast, and be willing to pay for good AI tokens so you can do them well and quickly. Push the projects you care about to 150%, and finish the ones you don't to a solid 80 in the least time possible. Don't waste time, so there's room left to learn more. Embrace new things eagerly, and be willing to spend money on the best AI.

4. Put half of your salary into VOO and QQQ as a regular monthly investment. You can also pick one or two tech companies you strongly believe in, like NVIDIA, Apple, or Anthropic once it goes public. Don't chase hot stocks, and don't speculate.

5. Find your partner earlier, rather than putting it off until you've been working for a few years. Watch how they handle problems and how patient they are. Look for the one who cares how you feel, stays optimistic, and stands on your side whatever happens.

6. Don't lend money to anyone, including relatives, classmates and old friends. Lending money will only weigh on you while you're young. Don't show off how much you earn, and don't spend carelessly. First grow your first pot of money to the point where you'd feel comfortable living for two years without working.

7. Keep learning, keep talking with people, and keep making fun things for the community, your circles and X. Build your GitHub and X accounts without chasing profit, so they grow into your personal brand. Always serve people assuming your users are good people.

8. Outside your main work, grow your own interests. Take some classes, whether photography, a musical instrument, sports, or even running and cycling. Life outside work won't get dull, and it gets you out and about more.

9. Health always comes first. Sleep matters, so never stay up late. Don't sit for too long, drink plenty of water, and walk more. Don't push too hard at work either. The first burst of energy fades, the second runs low, and the third is gone, so stretch out how long you can keep working.

10. In the end, find your own path to not having to clock in, as early as you can. You can still work, but you don't have to show up for a job. Earn money through your interests, expertise, strengths and what you keep learning, keep improving along the way, and enjoy making that money.

Tw93 · · 产品

🧑‍🚒 假如你的朋友需要发小红书,一定要把 Kami 推荐给他
最近 Kami 终于把我用了半年多的社交图功能给加上了,好内容值得好排版
只要告诉你的 Agent 说用 Kami 帮你把内容变成小红书卡片形式
那么就可以得到这种漂亮的效果了,甚至我还精挑细选了 8 种不同的底色
kami.tw93.fun/

假如小伙伴玩小红书,可以来看看我发的东西,哈哈哈“精致朋友圈”
xiaohongshu.com/user/profile/5c0c8d14000…

Tw93 · · 生活

疯狂推荐这个漱口水
这些年我大概买了快 30 瓶了
很好用,不辣,每次吃完饭用很清新

Tw93 · · 产品

One more thing, Mole CLI just landed on Linux, free and open source forever, as on Mac and Windows.

27 releases and 900+ commits in the past month. Linux went from first commit to 0.4.0 in four days, with the same menu, keys and safety rules as the Mac.

github.com/tw93/Mole

Tw93 · · 产品

都有 AI 了,还手写文档的小伙伴应该不多了,有一种邪不压正里面这年头“正经人谁写日记啊” 这种味道,后面可预计真正不借助 AI 写文章的场景会越来越少了。

其实我在 7 年前就开始想一个好用的 Markdown 笔记应用应该是怎么样的? 展开全文我认为首先是完全 md 语法规范,也支持社区的增强语法规范,但必须是规范的,这样 md 就可以在任何支持渲染的容器去使用,不需要绑定任何一个特定的软件,这也是 Markdown 和 Word 我认为最本质的区别。

同时应该是原生的、纯本地的、高性能的,不是套壳的,也应该是漂亮的,但不应该做成所见即所得,因为会很容易破坏原有 md 的纯净格式,当时用过为知笔记、有道笔记、Quiver、Bear、Obsidian 各种工具没有太顺手的,于是乎,当时就边学 Swift,在一个 2020 年初春节长假写下来了这个原生的 Swift Markdown 笔记应用,居然一直迭代到了现在,因为自己每天在使用,想到了一些东西,顺手的东西就全部给加上去了,这样越用越顺手。甚至后面为了防止笔记被改坏不好回去,写了一个类 git 的简单能力,可以帮你记录每次编辑改动,可以很好回到之前的版本,同时是一个本地功能,完全不上传任何云端。

终于在今年有空,把 iPhone 和 iPad 版本也支持了,当时为了搞通 Apple 开发者账号的付款居然最后写信给库克才解决,拿到证书后第一件事就去写这个啦,也因为有了它,我的所有开源的 Mac 相关的软件终于不弹危险未认证了,也给后来的 Mole Mac 版本创造了可能。

一直认为移动端的 Markdown 应用,重点应该是预览,我自己比较难在移动端上写一些长文,最多用于备忘记录,所以移动端的版本我更多做好渲染,以及阅读效果,尽可能做轻量一点,自己用了半年多,慢慢稳定了,可以推给大伙玩玩了。

到了 AI 年代,其实在妙言上的所有内容我都是手写的,想到啥写啥,特别是一些技术文章,包括最近的一些推特长文碎碎念,打字的顺畅感非常重要,不允许任何卡顿,预览也需要很舒服,排版需要更多是便于阅读,而非炫酷。

但以后用 AI 写文档的人会越来越多,所以我给妙言加上了 CLI 的功能,直接 miao open、miao new、miao search、miao list、miao cat 基本上可以覆盖笔记的新建、查找和读取,现在也够用了。

后面我想着 AI 时代的笔记应用应该是怎么样的?我想应该是阅读预览为主,继续轻量、美观、高性能,同时是 AI 友好的,以及是可以很方便的版本回滚的,最好是原生支持 git 的,这样 AI 用起来会很方便。

不过 AI 写出来的东西,其实有个 md 文件就够了,但最终还是给人看的,给人看的应该是成品,那还是需要人去读、去改,这样又回到了 Markdown 编辑器的范畴,只不过用的人会越来越少,我自己还是不想放弃独立琢磨写东西的能力,当然可以借助 AI,但不想把脑子全交给它,不然就没有创作的乐趣了。

假如小伙伴还有写东西的兴趣,可以去玩玩妙言 miaoyan.app/

Tw93 · · 思考

昨天有一个小伙伴私信问我,自己计算机专业,感觉现在 AI 可以完全代替他们的作业和论文,问我这个行业还有没有前景,以及对未来的一些建议。

我在想假如我回到大四的那一年,我会怎么做让自己未来之路少踩坑,走得顺一点,当然自己见识浅薄有限,欢迎交流。 展开全文

1. 肯定还是会选择计算机方向,然后选择自己最感兴趣的子方向,借助 AI 协作极大放大这一块的效果,好比一毕业就有一种带团队干活的感觉。

2. 选择北上深杭这样技术氛围好,科技行业多的城市,假如学历和技术足够厉害,就去北京清华以南北大以东学院路以西大钟寺以北这个区域内找一个合适的AI 公司工作,这样对于 AI 大模型的敏锐度会远高于其他的地方。

3. 工作时候,快速完成工作,愿意投资好的 AI Token 帮助自己快速做好事情,对于自己感兴趣的项目做到 150%的效果,对于不感兴趣的项目做到80分合格但是效率最高,不浪费时间,这样才有时间去学习更多的东西,非常积极的拥抱新事物,愿意花钱投资最好 AI 的使用。

4. 把工资的一半拿来投资 VOO、QQQ,当做定投,当然也可以选择自己极度看好的一两个科技公司,比如英伟达、苹果以及未来的要上市的 Anthropic,不碰任何 A 股、港股,不投机。

5. 更早的时候找到自己的另一半,不要拖到工作几年以后,多看另一半处理问题时候的能力和耐心,在意你的感受,乐观,出什么事情都站到你这边的那个。

6. 不要借钱任何人包括亲戚、同学、老乡,借钱会让年轻的你心情不好,不要炫耀自己赚了多少钱,也不乱花钱,先让自己的第一桶金能够到你感觉即使2年不工作也可以过得很舒服的地步。

7. 保持持续学习,持续交流,持续创造一些好玩的东西给到社区、圈子、推特,不逐利的去做自己的 GitHub、X 等账号,形成自己的个人品牌,永远抱着用户是好人的心态去提供服务。

8. 专业工作之外,培养自己的兴趣点,可以去上一些课,不管是摄影、乐器、运动甚至是跑步骑行这种,这样会让你工作之外,不会太无聊,也会让你可以多出去走走。

9. 身体健康永远放到第一位,睡眠很重要,一定不要熬夜,不要久坐,多喝水,多走走挺好的,以及工作不要干太猛,别一鼓作气再而衰,把自己可以工作的战线拉得更长一点。

10. 最终尽早找到自己让自己不上班的自由之路,可以工作但是可以不上班,通过自己的兴趣、专业、特长、学习去赚钱,并持续的在这个过程提高自己,很高兴的把钱赚了。

Tw93 · · 产品

With AI, writing code has gotten way faster, but verification is still slow, especially waiting on GitHub Actions CI. I spent this afternoon giving the CI for Mole, Kaku and Pake a full overhaul, and the results were great, so I want to share what I learned.…展开全文

Just copy the prompt below and hand it to your agent, and let's see how much time it saves you.

Speed up this repo's GitHub Actions as much as possible without dropping any necessary tests or checks.

1. Start by using gh to pull the last 50 runs and split each job into three phases: queue, setup and the actual test run. Look for patterns. Often the slow part is setup, not the tests themselves.
2. When a PR gets several pushes in a row, only run the latest one and cancel the older runs with cancel-in-progress. Don't do this on main, though: every commit on main needs a complete run before a release.
3. Commits that only touch the README or docs shouldn't trigger CI, so filter them out with paths. Watch out: a required check that gets skipped stays pending and blocks the PR from merging, so plan for that up front.
4. Prefer the built-in caching in setup actions: turn on cache in setup-node, setup-go has it on by default, and use Swatinem/rust-cache for Rust. Each repo only gets 10 GB of cache, so don't put anything in there you won't actually reuse.
5. Build twice in a row locally. The second build should take almost no time. If something still takes time, something is set up wrong, like a Dockerfile that copies the whole repo before installing dependencies, or a build script watching a file that doesn't exist. Check for anything similar.
6. Download prebuilt tools instead of compiling them on the spot with go install or cargo install, which eats a lot of time.
7. Run jobs that don't depend on each other in parallel, remove needs edges that don't pass any data, and split your longest test suite across multiple machines.
8. Free accounts can run at most 5 macOS jobs at once, shared across all workflows. Move anything that doesn't care about the OS, like formatting, linting and dependency audits, over to Ubuntu.
9. Don't do the same work twice, and don't do work you don't need. If another job already ran a check, don't run it again, and a job that only builds the frontend doesn't need to pull the whole Rust cache.
10. Once you're done, push and watch how the runs actually perform. Look at the timings, find the patterns, then do another pass to catch what you missed and get it as good as it can be. Finally, show me a before and after comparison.

Tw93 · · 生活

强迫症必备!
这款屏幕清理湿纸巾擦完电脑屏幕和刚拆封时候一样的感觉!

Tw93 · · 产品

One more thing, Mole is still simple, now it's also pro.

See which app eats your bandwidth, run a speed test, and hover any tag for Wi-Fi, disk and battery internals you'll never see unless you want them. Try it via Settings > Receive beta updates.

mole.fit/releases

0:29
Tw93 · · 产品

有了 AI 以后写代码速度快了非常多,但是校验的速度其实还是不快的,特别是在 GitHub 上跑 Actions CI 的速度,下午花了一些时间对 Mole、Kaku、Pake 的 CI 全部彻底优化了一下,效果很不错,分享给大伙。

你可以这样直接 Copy 下面给到你的 Agent,我们来看看可以帮你节约多少时间。 展开全文

帮我把这个仓库的 GitHub Actions 提速,在不减少必要的测试项以及功能校验下,最大幅度减少耗时。

1. 先使用 gh 拉最近 50 次的运行情况,把每一个 job 拆成排队、准备、真正跑测试的 3 个时间段,去看规律,可能很多时候慢的是准备的阶段,不是测试本身。
2. 同一个 PR 连着推好几次时,只跑最新的那次,旧的直接取消掉,用 cancel-in-progress 去实现。不过 main 分支不要这么做,需要确保发版 main 上的每一个提交都完整跑过一遍。
3. 只改 README 和文档的这种提交不要触发跑 CI,可以用 paths 过滤掉,另外被跳过的必需检查会一直卡住 PR 合不进去,这个要提前注意一下。
4. 优先用 setup 类 action 自带的缓存,setup-node 打开 cache,setup-go 默认就开,Rust 用 Swatinem/rust-cache,注意每个仓库缓存只有 10 GB,用不上的就不要放到这里面去。
5. 本地连着构建两次,第二次应该几乎不花时间,假如有多余的花时间的地方,一定是哪些写得有问题,比如 Dockerfile 装依赖前先 COPY 了整个仓库,或者构建脚本盯着一个不存在的文件,类似的问题记得检查。
6. 工具直接下载预编译好的,不要用 go install、cargo install 现场安装编译,这里会很费时间。
7. 不互相依赖的 job 并行跑,删掉不传数据的 needs,最长的那套测试拆到多台机器上去跑。
8. 免费账号的 macOS 最多同时跑 5 个,所有 workflow 共用,格式化、lint、依赖审计这些不挑系统的都移到 Ubuntu 去跑更好。
9. 做过的事情就不要做重复了,以及不要多此一举做事情,比如别的 job 跑过的检查就不要套重复再跑,比如只构建前端的 job 就不要多此一举去拉整个 Rust 缓存。
10. 改完去 push 那个上面测试运行效果,看运行的时间,发现规律,再次查漏补缺来优化一遍,做到最好,最后给到我优化前后的对比。

请 AI 机器号不要来刷评论,我的 Bot 每 15min 会自动检查来拉黑你们,此外有不小心误删的,非常抱歉,可以和我说。

Tw93 · · 生活

M Stand 这个咸芝士拿铁热的好喝!
小杯的量刚刚好,特别是第一口很不错!

Tw93 · · 工具

Sharing the 6 Chrome extensions I use every day.

1. Maple Bookmarks, made by me. I keep Chrome's bookmark bar hidden and open my bookmarks in a popup with a shortcut.
2. Maple NewTab, also made by me. A minimal new tab page that replaces Chrome's default and shows a beautiful…展开全文 painting every day.
3. Minimal Theme for Twitter / X. Highly recommended if you spend time on X. It strips away a lot of the clutter and makes X look much nicer.
4. PWA Install Block. I really dislike the install-app button in the address bar, and this hides it.
5. uBlock Origin Lite. The best ad blocker I've used, and the lightest on performance.
6. WebRTC Control. I care about privacy, and I don't want AI to know too much about me.

Tw93 · · 产品

Kami 1.20 is out. Your AI writes the words, Kami gives them good paper.

From the last three releases: charts drawn from data and checked before delivery, product sites in the product's own color, and Chinese and Korean pages that read well on iPhone.

github.com/tw93/kami

Tw93 · · 思考

来一波回忆杀,我的博客居然写了快 12 年了,从 2014 年底开始写,当时还是稚嫩的大学生,更有趣的是框架还是十几年前比较火的 Ruby 的Jekyll,几乎没有改过版本,当然也是我在写了 2 年 PHP 后感觉前端更有意思,学习 JS 的第一个项目。 展开全文

翻到了当时写了一篇 Vue、Angular、Avalon 框架的比较,当时 Vue 刚刚出来不久,Angular 被很多人开始喜欢,司徒正美老师的 Avalon 也有不少人用的那个时候,当时我还在用 Bootstrap 写全栈接商业项目赚生活费,看到知乎上的一个问题,花了一晚自习时间看官网文档写了几个demo,然后就写下了这篇很初级的对比文章,想到刚好可以沉淀到我的博客上,就随手提交了。

正是这个举动对我后续影响非常大,当时大学非常喜欢写代码,每天泡图书馆带着电脑按照书上的教程去学编程,然后想着让博客更加饱满,就会把学到的一些记录整理成文章搬到博客,差不多那两年时间把图书馆关于计算机基础、Web开发、编程思想的一些知名好书几乎都看完了,这一块对我帮助很大,以至后续工作的时候,感觉干活不是很难,就有更多时间来学习一些新东西。

转眼到了 2022 年,当时一年前学习 Swift 写了一个妙言 Markdown 笔记应用自己用,差不多成熟就开源了,后面想着要有一个发布更新日志的地方,顺手看到了这个小蓝鸟,非常好看,又顺手用谷歌邮箱一键登录了,发了第一个推特「妙言 - 轻灵的 Markdown 笔记本」,后面慢慢把妙言的更新日志以及做的过程就发到了推特,没多久被不少用妙言的小伙伴看到了,后续就交到了很多朋友,也是一个非常无意之举。

这个无意之举更有意思,我非常喜欢在推特发一些自己的想法和产品更新,当做一个完全没有顾忌的发言地方,一转眼居然玩了 4 年多了,正经推文差不多 2400 篇左右,由于推特自己的搜索和分类排序不好用,刚好需要查一些自己发的东西,于是我做了一个有意思的时候,让我的 Grok Bot 帮我的推特全部下下来了,也放了一份到我的博客上,我给这个栏目取名「片语」,开源是只言片语,每天会同步,这样也非常方便访问我博客的小伙伴可以看到一些轻松的片语。

说起写东西,我还有一个不变的习惯就是,不让 AI 代替我的思考、写作,所有大伙看到我的内容差不多都是手写的,虽然看着在 AI 时代很不潮流,但是我很喜欢这种输出,说不定这些文字可以成为以后硅基文明一个很重要的二十世纪中国年轻程序员的历史记录,哈哈哈,想想更有意思了,而且硅基人看着会感觉,呀和我们说话很不一样,是一个另类,得好好研究,这也是我在 AI 时代继续写一些看着没有什么用的东西的原因。

回到现实,我写博客没有走到大部分人走的那两条路,一年给自己博客换一个框架,装修一次,然后就是不写东西(没有喷的意思就是笑),也没有走另外一条,通过大量的 AI 生成补充很多看着非常丰富厉害的知识的模式(对不少人帮助很大),我的博客就只是我的一个 Markdown 渲染容器而已,表达的一个输出。

Tw93 · · AI

I spent an afternoon turning my AI coding habits into 20 rules that you and your agent can learn from together.

Briefing
1. Stop dictating the implementation. Describe the outcome you want. Models are good enough now, so aim high and don't worry that it can't be done.…展开全文
2. Give your agent the background and the goal: why it matters, who it's for, what great looks like. Then let go of the technical details.
3. Use the best model available. A weaker one just means more back-and-forth, wasted time and frustration.
4. Before adding a feature, setting or toggle, have the agent check with you first. Don't let it add features, toggles or copy nobody will use, and have it write your product rules down as it goes.
5. Think big, but have it build in the smallest usable steps and commit as you go, so you can roll back anytime.

Rules
6. Teach it how you talk: "look at" or "analyze" means read and report, "fix" or "change" means edit. Commits, releases and deletions always need your OK.
7. In AGENTS.md or CLAUDE.md, write only what it can't learn from the code, and what you've already ruled out. Never paste keys or tokens into the chat. If you do, treat them as leaked and rotate them.
8. Prompts alone won't hold hard rules. Enforce them with scripts, and only count it done when the checks actually pass.

Coding
9. Before writing code, have it explain what it will change and what that will affect, so you have the context to make the call.
10. Have it build a verification script for the project, run it at each checkpoint and show you the results. Over time, that becomes your agent's test suite.
11. Parallel agents each get their own worktree and commit only their own changes. Once everything is merged, run the full test suite.

Debugging
12. Find the cause before fixing. If two fixes fail, add logs and look at the real data instead of guessing.
13. If three fail, clear the context and start fresh with the problem, the goal and what you've already tried.

Review
14. Keep the agent that writes separate from the agent that reviews, so one doesn't bias the other.
15. Cap reviews at 3 rounds. From round two on, only review what's new.
16. The riskier the change, the deeper the check. Deletions, payments, and releases are one-way doors, so check them yourself and click through the real result.

Shipping
17. Every commit says what changed, how it was verified and how to roll it back. Your git history becomes great context for the next agent.
18. Use it yourself or give it to early testers first. Add a beta update switch so a bad release doesn't hit everyone.

Retro
19. Every production bug and user report gets a test case that runs before each release. Rerun them whenever you switch models to see if the new one holds up.
20. When fixing one thing, have it find and fix the same problem everywhere. Same for features and visual polish.

Tw93 · · 生活

今天化身做饭博主来教大伙如何整一个非常好吃的大闸蟹,如何蒸以及调灵魂料汁,哈哈这次我做得非常好吃。

首先假如家里有高压锅,一定一定要试试用高压锅来蒸,最好高压锅有一个隔板可以撑起来的那种,没有用正常的蒸锅也可以,下面放水上面隔空板子上放螃蟹,大闸蟹到了可以先泡水10分钟…展开全文,然后用刷子刷干净。

累计需要准备的配料,需要有两把葱花、一大块姜、2颗大蒜、买一小瓶蟹醋、几颗小青柠、2颗小米辣、生抽、白糖,这些都有用。

蒸螃蟹的时候,加适量冷水差不多距离板子还有1/3,冷水里面放两勺盐,让后放上支架,把葱切成段、姜切成片先摆一些在隔板上,然后大闸蟹背部朝下放好,每一个上面放一片姜,再弄一些葱段,然后淋一小圈料酒去腥,盖锅盖先大火等上汽,然后蒸 5min 左右,假如不是高压锅差不多需要 15-20min,然后闷3min的样子等高压锅气自然放完就可以摆盘了。

等高压锅的时候就可以来调料汁了,喜欢吃辣的可以切2-3颗小米辣,不喜欢吃辣的就不放也可以,然后切一点葱花碎、大蒜碎、姜碎,加一勺生抽、小半勺白糖,然后倒上蟹醋盖住你的这些料,最后最后,最关键来了,就是小青柠,你可以用刀切开一点,然后用手挤压汁到料碗里面,喜欢酸的可以挤3颗,不喜欢的可以挤一颗,加上这个会有很特别的好吃的酸味。

最后等出锅,热腾腾的,剪一根蟹腿捅出肉,蘸上灵魂料汁,吃一口真的非常爽,一定要试试高压锅+小青柠版本的大闸蟹。

Tw93 · · 产品

Mole tells you what your cleaned space would cost at Apple's storage upgrade price. I'm at $1,300 so far. Hover the Clean number at the bottom of the menu bar panel to see yours. How much have you saved?

Tw93 · · 产品

Kaku shipped 9 releases in October, mostly on remote access.

Run kaku remote and your Mac's terminals, running Claude Code sessions included, open in a browser on another device, and it now works with touch, on the web page and over remote desktop.

github.com/tw93/Kaku

Tw93 · · AI

假如把我的 AI Coding 经验沉淀成一些最佳实践规则,可以给到大伙复用,会是怎么样的,于是乎,今天一下午我就在琢磨这个事情,手写反复调整后给到大伙,可以让你和你的 Agent 一起去学习。

需求…展开全文
1. 从技术实现细节编程转变成许愿式编程,现在模型能力已经住足够强大,你只需要大胆把自己想法说出来,给它提足够高的要求和任务,不要担心 AI 实现不了。
2. 多给你的 Agent 讲做这件事情的背景和目标,少替他考虑技术细节,告诉它为什么要做,是给谁用的,做成什么样子你认为是最好的,具体的技术细节让它自己去判断,学会放手。
3. 优先用业界最好的模型,用差模型浪费的时候,一来回让你很不高兴的干活,二来也浪费了很多时间。
4. 要不要做一个新功能以及加一个新配置、开关的时候,让 Agent 来先问你找你确定,没有人用的功能、开关、文案都不要让他去加,慢慢把这一套你的产品规范让他沉淀起来。
5. 你的想法可以很大,但是让 Agent 干活的时候,尽量让他每次拆分成最小的可用功能版本,不要一次写非常多的代码,即使提交,这样出问题了你可以随时撤回。

规则
6. 让你的 Agent 了解你的说话风格,比如说看看、分析一下这个问题的时候,它只需要去读代码分析问题不要去改代码,你说修复、修改的时候才去动手,每次提交、发布和删除的时候都找你确定。
7. Agent md、CLAUDE md 里面除去重点的项目型要求外,只写代码里面看不出来的那种规则,以及之前被否决过的方案也可以分流进去。对你而言密钥和 token 不要贴给 Agent,贴了就当已经泄露了,立马去更换一个。
8. 对于Agent必须遵守的规则不要完全信 prompt,可能会拦不住,一定要用规则脚本去限制,让它真实跑一遍通过后才算完成了。

编码
9. 开始写代码之前,先让 Agent 讲清楚要改什么地方,会影响到什么功能和效果,这样是为了补充你的上下文,方便后面更好判断决策做什么。
10. 一定要让 Agent 基于项目补充一个能够让自己验证各种结果的脚本,每次特定的时候需要去跑一下验证效果,完成功能后需要检查,让他输出结果给你看,并基于此逐步沉淀你的项目 Agent 测试功能集。
11. 假如有多个 Agent 同时干活的时候,让他们各自用一份独立的 worktree 代码目录,同时提交的时候只提交自己改的代码,不要碰他人的代码,等到各自改完合到一起,再完整跑一遍测试。

排查
12. 修 Bug 的时候先找到原因再去动手,假如两次都没有修好,就去加一些日志看真实的数据以及运行效果,不要盲目去猜问题在哪儿。
13. 假如尝试修复了3次还不对,直接清空上下文重新写清楚你的问题、诉求,以及尝试过的解法让 Agent 重新开始。

审查
14. 让写代码的 Agent 和 Review 审查代码的 Agent 分开,保持解耦,这样一方不会被另一方带偏。
15. Review 代码的时候最多让 Agent 跑 3轮,防止跑着停不下来,第二轮开始只看新改动的地方,不要再把之前 Review 的再看一遍。
16. 对于你和 Agent 判断危险的改动,越危险让他查得越深,特别对于删除、付费、发布这种单向门改动你都需要自己去看一遍,点一下,看实际效果。

发布
17. 每次提交的时候,让Agent 写清楚做了什么,怎么去验证这个改动,除了问题怎么来回滚,用于给到后面开发过程中 git 记录本身就是 Agent 的一个非常好的上下文。
18. 需要对外的功能,可以先给自己用,或者给一批愿意尝鲜的用户使用,不要一上来就全量,可以在你的软件里面假如 beta 更新的开关,防止全量有问题,也是对用户负责。
18. 先给自己和愿意尝鲜的人用,没问题再发给所有人。

复盘
19. Agent 需要对每一个线上 Bug 和用户反馈的问题都留一个测试案例,每次上线前都来跑一遍,防止继续踩坑,包括你换模型的时候也可以把这些旧问题全部都跑一遍,看看模型是否还行。
20. 修复问题的时候,让 Agent 学会举一反三,把同类的问题一起修复掉。优化功能的功能,完善视觉的时候也是一样,防止只改一个地方,忘记了其他的地方。

请 AI 机器号不要再这个下面评论,我的 Bot 在自动拉黑你们,当然假如有误删,欢迎找我,非常抱歉。

Tw93 · · 产品

I started MiaoYan in 2020, and six years later it's on iPhone and iPad too. It's a native Markdown editor written in Swift, one App Store purchase covers Mac, iPhone and iPad with the same iCloud Drive folder, and the Mac app stays free and open source.
miaoyan.app/

更早的 Notes