ming

Building

[ming @mingwb]

My wife heard Obsidian pairs really well with LLMs, so I've spent the last couple days diving in to see how it could improve her workflow. But honestly? I can't get used to the editor. Why is there no WYSIWYG editing?

[ming @mingwb]

@dannypostma Hope you're doing well. I've watched your SEO course, and I thought it was really good.

[ming @mingwb]

@xicilion 没想起来用…

[ming @mingwb]

Some practices that have made my coding agents better to work with:

1. Add a setup script for new worktrees. Give each worktree its own environment file and database.

2. Review database schema changes and core type definitions carefully. Every line deserves review and discussion.

3. Use AGENTS.md as a map that points agents to more detailed instructions.

4. Define a strict implementation order for every feature: establish the frontend–backend contract first, implement the backend second, and build the frontend last. Document this workflow and make sure AGENTS.md points to it.

5. When the product direction is still unclear, ask the agent to build only the frontend with fully mocked data. Use it yourself for a while before committing to the product direction and backend implementation.

6. Add a VERIFICATION.md that explains exactly how agents should verify their work.

[ming @mingwb]

@naval I build products my child and I use every day.

[ming @mingwb]

Building websites and apps is becoming democratized. The SEO/GEO infrastructure around them shouldn’t be priced like a luxury.

[ming @mingwb]

@haydenbleasel My docs site: https://t.co/kZBAUxqXwu

[ming @mingwb]

Just set up my docs site with https://t.co/MIh5PLIqKh, built by @haydenbleasel.
I love tools with agent skills built in.

[ming @mingwb]

@nini_incrypto_ 可以花钱定目标。去花钱学 coursera 上的课程跟着上，每个月都扣钱；或者去考个证，有个 deadline。

[ming @mingwb]

Using AI to write documentation is much harder than using it to write code. I've been working on the documentation of Winda today, and what the AI produced is honestly just unreadable.

[ming @mingwb]

@xicilion 系统设计很难被训练到

[ming @mingwb]

I've been building Winda, and also using it every day to prep for the CFA — a finance exam that's basically thousands of pages you have to memorize.

Winda helps me understand the material and turn anything I select into flashcards while I'm reading. Then spaced repetition makes sure I remember it.

https://t.co/OnsNa3qna5

[ming @mingwb]

@TanzilaSha9574 Create flashcards from anywhere you read.
https://t.co/nHMNTJU5V9

[ming @mingwb]

结构化数据才能有更丰富的交互。

[ming @mingwb]

结构化这条非常认同。结构化的数据才能作为资产来管理，机会会多一点。pdf word 等待也是结构化数据，只不过市场太大，在模型的主赛道上。

[赵纯想 @chunxiangai]

laper 方法：

1、离大厂三条街。

2、找一个结构化非常复杂的领域。

譬如剧本，有极为复杂的结构化要求，CodeX能写剧本，但无法输出统一好莱坞格式剧本，没有版本管理 + 选角表 + 堪景表 + 分场表 + 分镜表。

事物足够复杂，不会被大模型桌面端应用吃掉。

3、除了剧本之外，类似的泥潭般的结构化要求领域，包括但不限于：工程标书，演艺策划，考卷出题。

4、找到过去几十年里深陷数据结构泥潭的用户，在他们手边装一个 Agent，将泥潭变成荷塘月色（UIUX）。

5、虽是荷塘月色，但独赏只有凄凉。

——必须有多人协作场景。没有多人协作场景的，没有协同需求，则没有技术难度。而 CRDT，AI Coder 一时半会儿玩儿不明白。

6、屏蔽创业者播客。屏蔽融资新闻。在泥潭里捞出你的第一个付费用户。

7、此为纯想扭屁股论之后的纯想泥潭论。AI应用创业，目前是绝对冰点。没有比这更好的时机。

[ming @mingwb]

@sailfishcc1 claude code 的记忆我觉得还行

[ming @mingwb]

为什么 Codex、Cloud Code 的 sandbox 环境，都没有用 Docker 的方式安装依赖呢？

[ming @mingwb]

一天 5 个 PR，一个 PR 可能有 2 到 3 次 build。在 Vercel 上面，没有任何流量的话，可能一个月都要花将近 20 美金左右...