Bimon.
Work Writing About

Hi, I’m Bimon.
I worked in sales for 19 years. Now I build tools and systems for people and agents.
你好,我是 Bimon。
我做了 19 年销售。现在也在做给人和 agent 用的工具和系统。

This is the whole company right now: no team, no funding, one laptop. Everything below runs on a machine you own — the transcripts, the notes, the words you look up. 这一页就是当前的全部,没有团队,没有融资,一台笔记本。下面这些东西都跑在你自己的机器上——转写、笔记、查过的词。

Now这个月 · · learning how to make video在学怎么做视频


Work做的东西 six things · all running六件 · 都在跑
AV2Text 0.1.16

Turns audio, video and the text inside images into a transcript you can use as it is. Fully local, runs offline: your original recordings and images never go to anyone's cloud. 把音频、视频和图片上的文字,变成一份直接能用的文字稿。完全本地,离线可以运行;音视频原件和图片原件不会上传到云端。

Personal RAG desktop · MCP桌面 · MCP

Personal RAG turns a folder of notes on your computer into a private knowledge base you can question. Personal RAG 把你电脑上的笔记文件夹,变成能问答的私人知识库。

Ask it something and it answers only from what you have written, marking the source of every sentence. When there is no evidence, it says it does not know rather than inventing one. 问一句话,它只从你写过的内容里找答案,并标出每一句的出处;找不到依据时,它会直接说不知道,而不是编一个。

PDFs, Word files and web pages dropped into the inbox become searchable notes. It also plugs into assistants like Claude and Cursor, so they answer from your notes. Everything stays on your computer — no upload, no account. PDF、Word、网页丢进收件箱,就会自动变成可检索的笔记;也可以接到 Claude、Cursor 这类 AI 助手上,让它们用你的笔记回答。内容都留在你自己的电脑上,不上传,也不需要注册账号。

Lexi 0.1.5

Lexi explains more than single words: phrases, sentences, entire paragraphs. Lexi 能解释的不光是词语,还有词组、句子,甚至段落。

What it is really for is crossing into another field quickly — code an agent wrote, a senior meeting compressed into political language, or the industry jargon buried in a meeting transcript. 它最重要的功能,其实是帮人快速理解跨领域的文本:agent 写的代码、一场高层会议(往往充满凝练的政治用语),以及音频会议转写结果里那些行业黑话。

Of everything on this page, Lexi is the only tool that helps a person keep pace with their agent — so that talking to it does not leave you permanently the fool. 在我所有的作品里,Lexi 是唯一一个帮助人类高效对齐 agent 进度的工具——让你在跟 agent 交流时,不至于永远像个傻子。

Quality Console质量数据看板 factory · daily工厂 · 每天

Eighteen cloud functions, six inspection forms for six roles. In daily use inside a factory — it is, plainly, a work-order reporting system. Phones and computers both sign in. 18 个云函数、6 个岗位 6 张检验表单,工厂里每天在用。其实就是一个工单提报系统,手机、电脑都可以登录。

dl 2026.10.23 · 19 cmd · 23 mcp

A downloader with no window. One command, one JSON on stdout, 29 tests, exit codes 0/1/2/130. 没有窗口的下载器。一条命令,stdout 上一个 JSON,29 个测试,退出码 0/1/2/130。

Cobuild 0.1.0 · for myself给自己

A calendar I wrote for myself, because my agent and I kept two sets of the same tasks. Its interface borrows from Notion's calendar; the backend and the provenance column are mine. 给自己写的日历——因为 agent 和我各自记着一套一样的任务。界面借鉴了 Notion 日历,后端和「来源」那一列是我自己的。

Hand it straight to your agent. Nothing more to say. 直接把它交给你的 agent 吧,没什么好说的。


Writing写的东西 01 piece一 篇

One year, from no code to six tools一年,从不会写代码到六个工具

2026 · 10

A nineteen-year salesman, OpenClaw the morning after it blew up, and why I believe the cost of copying capability is now close to zero. 一个做了十九年销售的人,在 OpenClaw 爆火的第二天早上就装上了它;以及我为什么相信「能力复制的成本已经接近 0」。

read it读全文

1 · By the time I woke up, everyone had it installed一 · 我醒来的时候,全世界都装好了

I live in China. OpenClaw went viral on a US timeline, which means it happened while I was asleep. The next day I had it running on my own machine. It carried me through the phase every beginner goes through — where every idea is exciting and none of them work yet. After that came WorkBuddy, OpenCode, Claude Code, Codex. Today I use DSH, and I like it.我在中国。OpenClaw 是在美国的时间线上爆的,也就是我睡觉的时候。第二天我就在自己电脑上装好了。它陪我度过了每个初学者都会经历的那一段:每个想法都兴奋,但没一个跑得通。之后又陆续用了 WorkBuddy、OpenCode、Claude Code、Codex,现在用的是 DSH,我感觉蛮好的。

2 · The sentence I keep quoting二 · 我一直引用的那句话

Lei Jun said that the internet drove the cost of copying information to almost zero, and that AI drives the cost of copying capability to almost zero. My year is a footnote to that sentence. A year ago I could not program at all. It did not stop me from finishing the things on this page — and the pace kept accelerating as the models got better. The last few products each took under two months.雷军说过:互联网时代,把信息复制的成本几乎打到了 0;AI 时代,把能力复制的成本几乎打到 0。我的这一年就是这句话的注脚。一年之前我完全不懂编程,这并不影响我做出这一页上的东西;而且效率是随着 agent 和模型的发展飞速提高的,最近几个产品都是两个月之内做完的。

3 · AV2Text took half a year三 · AV2Text 花了半年

It was the slow one. I started in Python, was unhappy with the speed, and rewrote the whole thing in Rust. Then months of comparing ASR models, adding and cutting features, redesigning the interface, until it settled into the version you can download today. It will keep moving.它是最慢的一个。一开始用 Python 写,对性能和速度不满意,后来用 Rust 重写了。接着是对比各种 ASR 模型、增删功能、改 UI,才稳定到你现在能下载的这一版。后续还会不断迭代。


About关于 oct 2025 →

I met vibe coding in October 2025. On 31 January 2026 I installed OpenClaw — early enough to be in the first wave — then OpenCode, then the ones everyone knows: Cursor, Windsurf, Claude Code, Codex. Today I use DSH. I started by writing browser extensions to fix OpenClaw's multimodal input, then moved into projects: some abandoned, some kept chasing the models as they moved. Most of that time went into ASR and TTS, and I have opinions about both. 2025 年 10 月接触 vibe coding;2026 年 1 月 31 日装上 OpenClaw,应该是全球比较早的一批,不久之后接触 OpenCode,然后是大家都熟悉的 Cursor、Windsurf、Claude Code、Codex,现在用 DSH。最早做了一些浏览器插件,解决 OpenClaw 的多模态交互问题;后期慢慢接触一些项目,有中途放弃的,也有跟随模型的发展不断跟进的。接触 ASR 和 TTS 领域多一些,也有自己的一些想法。

What survived is on this page. The console was built for a manufacturing company; almost everything else is built to be called by an agent — dl was written almost entirely for one. 留下来的就是这一页上的东西。工单系统是我为一家制造业企业开发的;除它之外,我所有的工具都是 agent 友好的,甚至 dl 几乎是专门给 agent 写的。

I think the entity that does the work is going to be an agent, not a person. So while I build agent tools, I also try to keep the human in step: that is why Lexi exists. 我认为未来具体工作的承载主体必然是 agent,而不是人。所以在开发 agent 工具的同时,我尽量让人类跟上 agent 的节奏——这就是 Lexi 的由来。

The pair I believe in most is AV2Text + Personal RAG, because it grows your agent's private knowledge fast — you still supply the taste. Public information is already inside the models; the private kind sits in long videos on YouTube and Bilibili. So when you find a creator you like, well: you understand what to do. 而 AV2Text 和 Personal RAG 的组合,是我当前最看好的:它可以迅速扩大你 agent 的私域知识库,当然审美还得你来把关。众所周知,公域信息 AI 都了解,私域信息大量沉淀在 YouTube 和 B 站这些中长视频网站上。有了这两件,遇到喜欢的博主时——话都说到这份上了,你会懂的。

The loop这一圈 AV2Text → Personal RAG → agent

Diagram: AV2Text keeps feeding transcripts into Personal RAG, whose three cells fill up one by one. The agent — a small heart — keeps drawing from it, and the three stars above it brighten as it goes.图示:AV2Text 不断把转写喂进 Personal RAG,它的三格电量逐格填满;你的 agent(一颗会长眼的心)不断从里面取资料,头顶的三颗小星星随之变亮。

AV2Text Personal RAG your agent你的 agent