Lewis21

AI is easy to use. Hard to use well.
Exploring human–AI collaboration, workflows, and AI at work.
Head of AI Operations at a startup.

[Lewis21 @time_points]

A decade-plus of Chrome bookmarks. 1000+ of them. And the sites I saved? Still unfindable.

Bookmarks don't scale — past a certain point they become a pool of dead water. Everything's "in there," but if you can't dig it out, it's the same as nothing.
So I built myself a Chrome extension with Claude Opus 5.5. Core features:

1. Auto-indexes all bookmarks (existing + new — you can also add tags manually), with natural-language search. "Which image-compression sites did I bookmark?" → straight to the answer, no digging through three layers of folders.

2. Suggests deletions based on open frequency: bookmarks untouched for ages get listed for one-click cleanup.

3. BYO API key for both indexing and AI search. I plugged in the Mimo V2.6 Flash model — less than $0.01 per search.

If you've ever drowned in bookmarks, like or leave a comment — I'll open-source the project soon.

[Lewis21 @time_points]

攒了十几年 Chrome 网站书签，1000+ 个，最后发现：收藏的网站却仍然找不到。

书签一多就变成一潭死水——资源明明就在里面，但翻不到，等于没有。

痛定思痛之下，用 Claude Opus 5.5 给自己写了个 Chrome 插件，核心功能如下：

1. 给全部书签（存量+新增）自动建索引（也可以自己手动添加标签），支持自然语言搜索。"收藏的图片压缩网站有哪些？"，直接搜出来，不用翻三层文件夹；

2. 基于打开频率给删除建议：长期没打开的书签列出来，一键清理；

3. 索引和 AI 搜索都支持自带 API key。我接入的是 Mimo V2.6 Flash模型，单次搜索成本不到几分钱。

如果你也被书签淹没过，欢迎点赞或留言，我尽快把这个项目开源。

[Lewis21 @time_points]

攒了十几年 Chrome 网站书签，1000+ 个，最后发现：收藏的网站却仍然找不到。

书签一多就变成一潭死水——资源明明就在里面，但翻不到，等于没有。

痛定思痛之下，用 Claude Opus 5.5 给自己写了个 Chrome 插件，核心功能如下：

1. 给全部书签（存量+新增）自动建索引（也可以自己手动添加标签），支持自然语言搜索。"收藏的图片压缩网站有哪些？"，直接搜出来，不用翻三层文件夹；

2. 基于打开频率给删除建议：长期没打开的书签列出来，一键清理；

3. 索引和 AI 搜索都支持自带 API key。我接入的是 Mimo V2.6 Flash模型，单次搜索成本不到几分钱。

如果你也被书签淹没过，欢迎点赞或留言，我尽快把这个项目开源。

[Lewis21 @time_points]

用huashu的skill+opus5.5把最近看到的一篇微信公众号文章转成了一段科普视频，效果真的很不错！

[花叔 @AlchainHust]

Huashu-art-motion发布12小时后，已经500 stars🌟了。

等超过5000+星标的时候，我再猛猛更新一波吧！

会大幅度增加视觉风格，以及让GPT、国产顶流模型也能做出类似效果，敬请期待～

[Lewis21 @time_points]

已经把这段提示词加入ChatGPT的自定义指令中：【当我要求你进行解释时，通常情况下请用接近 ASD-STE100 的清晰写作原则解释，但不要机械遵守规范。控制术语密度，少用长句和从句，一个句子尽量只表达一个核心关系。不要为了避免重复而替换同义词。明确写出动作主体和因果关系。先建立 mental model，再补充专业术语。必要时使用具体例子和类比。】

[Andrej Karpathy @karpathy]

We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:

Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:

Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:

Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:

Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!

In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.

[Lewis21 @time_points]

If your CEO hired someone to "own" org development, your org development is already dead.
OD without power is theater: surveys, workshops, culture decks. The real levers — headcount, incentives, reporting lines — only the CEO can pull.
A CEO can delegate the craft. They cannot delegate the accountability.

[Lewis21 @time_points]

The AI assistant I've been using — Meta's Muse (I call it Milo) — is honestly one of the best I've tried. It handles my inbox, manages my calendar, and gets things done for me like a real personal assistant.

Join me 👇
🔗 https://t.co/ZfWVxDuMW9
🔑 Invite code: [45VO58]
We each get 1 billion tokens when you join. Win-win!

[Lewis21 @time_points]

推荐两个不错的workbuddy学习教程：
1. 《Workbuddy蓝皮书》，由苍何、甲木未来派、摸鱼小李、袋鼠帝、刘聪NLP等一众大佬开发而成，一本以真实任务为主线的 WorkBuddy 中文使用手册：https://t.co/ogJyWTZEVY
2. 《WorkBuddy 实战绿皮书》，由HR领域AI专家“我是雷神”开发，带你把 AI 真正落进你所在的行业：https://t.co/Jd07EdvWUW

[Lewis21 @time_points]

Most bad AI work isn't a model problem. It's a delivery-standard problem. You know what "good" looks like. The AI doesn't. Here's the three-step fix that actually works.

1. Sharpen before you assign.
Don't drop the full task on day one. Talk it through first. Ask for rough samples, a simple cover, a rough page, anything cheap. Round by round, pull the unspoken standards out of your head and into the chat.

2. Write it down, then run it clean.
Once you agree, turn that chat into one clear brief. Open a fresh window. Hand the whole document over once. Don't keep patching mid-run and waiting through another long loop.

3. Keep the brief. Reuse it.
That document isn't just for this job. It's team knowledge. Next similar task, start from the brief, not from zero. Quality goes up. Cycle time goes down.

Collaboration engineering beats better prompts.

[Lewis21 @time_points]

Codex hit me with "Selected model is at capacity." All my models went dark. My projects, sessions, and harness configs all live there, so leaving Codex wasn't an option.

I stayed. I just made Codex talk to third-party models too, with Open Codex:
https://t.co/Lb7XI874Y4

How: I didn't install it by hand. I gave Claude Code this prompt and let it do the work:

"Help me install this plugin: https://t.co/Lb7XI874Y4. It lets Codex use third-party models while still using OpenAI's own models. First check whether installing it would wipe my Codex sessions. If it would, warn me before installing. Back up my Codex config before install so I can restore it if something breaks."

That one prompt got me out of capacity hell without abandoning the harness I already built.

#codex #ClaudeCode

[Lewis21 @time_points]

How I pick GPT-6 Astra reasoning levels, without lighting my quota on fire.

Skip Ultra. It costs a lot and the extra juice is tiny.

Non-coding and simple coding: low or medium.
Hard coding: high or xhigh is enough.

A real bone to chew: leave Codex. Use ChatGPT 6 Pro, plug in the official GitHub plugin, let it open the PR. The $200/mo ChatGPT Pro 20x plan gives you 200 of those a week.

The combo that actually works:

Easy job: gpt-6 low or medium.

Harder job: start on high, with plan mode. Let it audit, research, and write the plan.

Then drop to low or medium and execute.

Planning is where you want it to think. Once the plan is on the table, the rest is just doing the work. Lower effort is fine there.

#gpt6

[Lewis21 @time_points]

3. Session-to-session, I write five lines and stop. What is in flight. Why it matters this week. What the next session is for, in one sentence. Where the spec, plan, and diff already live. What I believe but have not verified. A guess written as a fact becomes the next agent's floor.

[Lewis21 @time_points]

I run multiple agents every day. The hard part isn't the model.
It's the handoff.
Agent A finishes.
Agent B starts from a worse brief than a junior hire would get.
Context dies in the gap.
That's collaboration engineering. Prompting is just the surface.

[Lewis21 @time_points]

2. A new chat that starts with "continue" is not a handoff. It's a memory leak. The last session already paid for the context. The next one should inherit the live thread, the reason it exists, and pointers to what is already written down. Not a vibe. Not the whole transcript.

[Lewis21 @time_points]

In the AI era, prompt engineering may only be a surface-level skill. The deeper skill is collaboration engineering.

How we think about AI shapes how we use it.

If we treat AI as a traditional tool, we naturally expect a simple transaction: I give an instruction; it delivers a result. If the result isn’t good enough, we conclude that the tool isn’t good enough.

But if we treat AI as a cognitive collaborator, the way we work with it changes.

We start asking ourselves:

Did I make the goal clear?
Can I give useful feedback?
Am I willing to let it challenge my assumptions and point out where I might be wrong?

Perhaps the key to using AI well isn’t learning more prompts. It’s learning how to become a better collaborator.

A good collaborator can define goals, give feedback, and be challenged.

And great human–AI collaboration isn’t just about humans constantly correcting AI.

It works both ways:

We help AI understand the problem better.
AI helps us understand the problem differently.

That’s when AI stops being a tool that simply follows instructions and becomes a cognitive partner—one that can think with us, iterate with us, challenge us, and help us move toward better answers.

[Lewis21 @time_points]

https://t.co/hbdIcmwJyY

[Lewis21 @time_points]

The idea is cool, but could you please write a more understandable tutorial? I downloaded this desktop app, but haven't figure it out how to use it. The 'Memories' listed my recent sessions from other local agents, but all are labeled 'Processing Failed', so not sure if it actually works.

[Lewis21 @time_points]

🤖 日常跟多个Agent协作已成常态，但Agent间的配合是老大难。本周发现两个解法，底层逻辑截然不同： ① Raft（https://t.co/P2OotePxaJ）：自己创建Agent团队，非接入外部工具。每个Agent通过本地daemon运行，拥有独立身份和持久记忆，能认领任务、彼此交接、互相review。Claude/Codex/Hermes任选底层引擎。你的电脑就是Agent运行环境，人在指挥位决策。Free起，Pro $8.80/月。

② Bloome（https://t.co/0ruFmDvQIg）：把已有的本地Agent（Claude Code/Codex等）和云端模型（Claude/GPT/DeepSeek）拉进同一个群聊。多个Agent在对话中互相校验打磨，支持多模型同屏对比。Agent还能生成Widget可视化输出。目前免费内测。 Raft是「建团队」，Bloome是「组局」。
#AI协作 #Agent #效率工具

[konmari @konmari_tweet]

キラキラ水面が本物みたいに揺れるWebアプリ作っちゃいました！
Claude Fable 5の作例でいろんな人の美しい作品を見てたら、 「これ、自然な表現めっちゃ上手いなぁ✨」となって、 勢いでインタラクティブな水面リップルアプリを作ってみた。
マウスで撫でると波紋が広がって、 花びらがゆらゆら動きます🌸
これがHTML？画面にあるのが不思議な感覚。
※背景画像はGPT Image 2で生成しました。

[Lewis21 @time_points]

使用ai agent查询本地微信消息对于加了很多微信群以及维护微信群的人来说都是刚需，今天刚好又有人问到我该怎么操作，就把这个方法分享给大家：
1. 安装必备神器：https://t.co/76GYMiBcfV
2. 结合你用得顺手的本地agent工具（claude code, codex, workbuddy都可以），按照下图操作即可（windows亲测有效，mac尚未测试）。