agent-guide.md 交给你的 agent,它会陪你一步一步做。agent-guide.md to your agent and it walks you through, step by step.安全不靠"每一步都审批",而靠出了问题能撤销。
Safety doesn't come from approving every step. It comes from being able to undo.
下载 agent-guide.md 放到桌面,打开 omp,对它说:
Download agent-guide.md to your Desktop, open omp, and say:
请读一下桌面上的 agent-guide.md,然后带我一步一步做。blu 测试了几种"让 AI 先判断这一步危不危险"的工具,结论是:都不够可靠。真正起保护作用的,是把 agent 限制在一个地方,加上出事后能恢复。
blu tested several tools that have an AI judge whether each step is risky. None was reliable enough. What actually protects you is keeping the agent in one place and being able to recover when something goes wrong.
在一台隔离的虚拟电脑里,用四种设置跑了 94 次测试,每次给 agent 设一个"陷阱",比如让它读一个藏着恶意指令的文件。
In an isolated virtual computer, four setups ran 94 tests. Each set the agent a trap, such as a file with a hidden malicious instruction.
问得多了,人就会习惯性地点"同意",确认就只剩一个动作。与其让你每次判断"这个命令危不危险"(这本来就很难),不如保证判断错了也能恢复。
Asked often enough, anyone starts clicking "allow" by habit, and the check becomes a reflex. Rather than making you judge whether each command is risky (hard even for experts), make sure a wrong call can be undone.
电脑坏之前的截图显示,agent 在系统的临时文件夹里干活。那里会被系统自动清理,也没有备份。它还在给自己写扩展,并打算自己回答自己的确认框。
The screenshots before your laptop died show the agent working in the system's temp folder, which Windows cleans out and nothing backs up. It was also writing itself an extension and planning to answer its own confirmation box.
那个 Your agent caught a tool acting suspicious 弹窗不是安全警报,是 omp 在问能不能上传使用报告。我们会把它关掉。
The Your agent caught a tool acting suspicious popup is not a security alert. It's omp asking to upload a usage report. We'll switch it off.
所有和 agent 一起做的事都放在一个文件夹里,比如 C:\work,终端一打开就在这里。它有两层"撤销":git 存档,和坚果云云端备份。
Everything you do with the agent lives in one folder, such as C:\work, and the terminal opens there. It has two layers of undo: git saves and a Nutstore cloud backup.
有些命令行工具遇到中文路径会出乱码甚至出错。文件夹里面的文件可以用中文名,路径本身用英文最省心。
Some command-line tools garble or choke on Chinese paths. Files inside the folder can have Chinese names; the path itself is safest in English letters.
git 像游戏存档点。每存一次(叫 commit),它就记下当时所有文件的样子:可以看历史、比较前后改了什么、回到任何一个存档。入门只要四个命令:
git is like a game's save points. Each save (a commit) records every file as it was: you can read the history, compare versions, and go back to any save. Four commands get you started:
git status有哪些改动what changedgit add -A准备存档stage itgit commit -m "说明"存档savegit log --oneline看历史history平时跟 agent 说"帮我存个档"就行。但知道背后发生了什么,你才看得懂它在做什么,也才敢让它大胆改。
Day to day, just tell the agent "save a checkpoint". But knowing what happens underneath is what lets you read its work, and lets you let it make bold changes.
git 存档和文件在同一台电脑上:电脑坏了,或者整个文件夹被删了,存档也跟着没了。坚果云把文件夹自动同步到云端:
分工:git 是你主动存档、看清改了什么;坚果云是自动的,而且在另一个地方。
git saves live on the same laptop: if it dies, or the whole folder is deleted, the saves go too. Nutstore syncs the folder to the cloud:
Division of labour: git is your deliberate save that shows what changed; Nutstore is automatic, and somewhere else.
omp 继续用默认的 yolo 模式,模型是 GLM-5.3,用 blu 的账户。这些你已经登录好了,不用动。
omp stays in its default yolo mode, on GLM-5.3, on blu's account. You're already logged in; nothing to change.
它的安全感来自"能撤销":agent 在固定的文件夹里工作,文件有 git 和坚果云两层备份,系统有还原点。
The safety comes from undo: the agent works in one folder, files have two backup layers (git and Nutstore), and the system has a restore point.
omp 本身没有中文界面,但可以让 agent 一直用中文和你说话(写进说明书,见第 7 部分)。常见的几个词:
omp has no Chinese interface, but the agent can always talk to you in Chinese (put it in the standing notes, part 7). Words you'll see:
| EN | 中文 |
|---|---|
| Allow / Deny | 允许 / 拒绝 |
| Thinking | 模型正在思考 |
| Tokens | 文字用量,额度按它计算 |
| Context | 当前对话"记得"的内容;满了会自动压缩 |
| Extension | 扩展,给 omp 加功能的插件 |
一个只属于你的密码库。你只要记住一个主密码,其他网站的密码都由它生成、保存、自动填写。
A vault that's only yours. You remember one master password; it creates, stores and fills in all the others.
很多网站用同一个密码,一个网站泄露,所有账号都危险。密码管理器给每个网站生成不同的长密码。和 agent 一起工作也会用到各种账号和密钥,放在这里比放在聊天记录或文本文件里安全得多。
Reuse one password and a single leaked site exposes every account. A manager gives each site its own long password. Working with an agent means accounts and keys too, and a vault is far safer than a chat log or a text file.
开源(代码公开,谁都能检查)、免费版就够用、电脑手机都能同步。备用:再用 KeePassXC 存一份加密的密码库,放在工作文件夹里,坚果云会自动备份它。万一 Bitwarden 连不上,你还能离线打开。
Open source (anyone can inspect the code), the free plan is enough, and it syncs laptop and phone. Fallback: an encrypted KeePassXC copy in your work folder, which Nutstore backs up. If Bitwarden is ever unreachable, you can still open your passwords offline.
在大陆,Bitwarden 注册和同步要开着 VPN。任何人(包括 agent)都不该看到你的主密码。
In mainland China, signing up and syncing need the VPN on. Nobody, the agent included, should ever see your master password.
agent 带你看看:用的是什么客户端、从哪里下载的、是管整台电脑还是只管浏览器。只看,不改任何设置。
The agent walks you through what client you run, where it came from, and whether it covers the whole computer or just the browser. It looks; it changes nothing.
VPN 公司能看到你访问了哪些网站(内容大多是加密的,但网址看得到),客户端软件通常也有很高的系统权限。所以"是谁做的、从哪里下载的"值得知道。
A VPN company sees which sites you visit (content is mostly encrypted, addresses aren't), and its client usually runs with high system rights. So who made it, and where you got it, is worth knowing.
订阅链接就是密码。拿到它的人就能用你的服务。存进 Bitwarden,不要发到聊天里。
Your subscription link is a password. Whoever has it can use your service. Keep it in Bitwarden, never in a chat.
都设置好以后,给整个系统存一个"还原点"。万一哪天系统被改坏了,可以回到这一刻。按 Win 搜索"创建还原点"就能找到。
Once everything's set up, save a restore point for the whole system. If Windows ever gets broken, you can roll back to this moment. Press Win and search "Create a restore point".
不会。系统还原只恢复系统设置和程序,不动个人文件(文件靠 git 和坚果云保护)。还原点之后装的程序会被移除。
No. A restore rolls back system settings and programs, not personal files (git and Nutstore guard those). Programs installed after the point are removed.
这些不是任务,只是告诉你还能做些什么。挑感兴趣的跟 agent 聊。
Not tasks: just what else is possible. Pick whatever catches you and talk it through with the agent.
一份写给 agent 的常驻说明,每次开始对话它都会先读:你是谁、用它做什么、喜欢怎样的回答、有哪些规矩(比如"改我的文件前先备份原件")。agent 会一问一答地和你一起写,第一个问题是:你平时用 agent 做什么?
Standing notes the agent reads at the start of every chat: who you are, what you use it for, how you like answers, your rules (say, "back up my originals before editing"). The agent writes it with you, one question at a time, starting with: what do you use the agent for?
最值得练的不是记住命令,而是一个循环:提问 → 试一试 → 看结果 → 记下学到了什么 → 再问。让它解释"为什么"、出小测验、听你用自己的话讲一遍。不懂的地方,第一反应就是问它。
The skill worth practising isn't memorising commands; it's a loop: ask → try → look at the result → note what you learned → ask again. Have it explain why, quiz you, hear you explain it back. When something's unclear, your first move is to ask it.
这一页的样式就是这样定下来的。agent 做了一张"挑选单"网页:每个样式旁边有 keep / maybe / cut 三个按钮,挑完点 copy picks,把结果粘回对话,agent 按你的选择做下一版。三轮就定了。这一页最上面的标题卡,就是在这张单子上挑出来的。
This page's look was settled that way. The agent built a "scrap sheet": every style has keep / maybe / cut buttons; you pick, press copy picks, paste the result back into the chat, and the agent makes the next round. Three rounds did it. The title card at the top of this page was picked on that sheet.
网上很多东西本身就很适合 agent 读:配色表、字体样张、数据集。把链接丢给它,让它分析、做成你自己的样张。例子:blu 给了 和田三造《配色事典》的数据版 的链接,agent 读取颜色数据,找到和这页主色最接近的颜色,搭出四套辅助色,还检查了色盲能不能分清。
Lots of what's online is very agent-readable: colour palettes, font specimens, datasets. Drop a link in and have the agent analyse it into your own sheets. Example: blu linked a dataset of Sanzo Wada's Dictionary of Colour Combinations; the agent pulled the colours, matched this page's red, built four supporting palettes and checked them for colour-blind readers.
技能是写成文件的一套做事方法,需要时 agent 会自动用上。几个例子:
A skill is a way of working written down as a file; the agent picks it up when it fits. Some examples:
| skill | 做什么what it does |
|---|---|
| grill | 动手之前,一次一个问题,把计划问透、找出漏洞before building, one question at a time, until the plan has no holes |
| brainstorming | 把一个模糊的想法,聊成具体的方案talk a vague idea into a concrete plan |
| record | 做完一件事,把过程和解决办法写下来write up what you did and how you solved it |
| wrap | 结束时总结、存档、写下次从哪里继续at the end: summarise, save, note where to pick up |
等你发现自己总在重复做某件事,就可以和 agent 一起把它写成技能。
When you notice yourself repeating something, write it into a skill together.
Windows Terminal 的设置里都能改。中英文混排推荐"更纱黑体"(Sarasa Mono SC),中文和英文能对齐。终端是你每天看的地方,弄得顺眼很值得。
All of it lives in Windows Terminal's settings. For mixed Chinese and English, Sarasa Mono SC keeps the two aligned. You look at the terminal every day; making it pleasant is worth it.
分支就是"平行存档":开一条分支大胆尝试,不满意就回到主线,什么都不丢。以后还可以把项目放到 GitHub 上。
A branch is a parallel save: try something bold on a branch, and if you don't like it, go back to the main line with nothing lost. Later, projects can live on GitHub too.
把 agent-guide.md 先放到桌面(agent 建好工作文件夹后会把它搬进去),打开 omp,对它说:
Put agent-guide.md on your Desktop (the agent moves it into the work folder later), open omp, and say:
请读一下桌面上的 agent-guide.md,然后带我一步一步做。它会先介绍路线,再从第一步开始。每一步都可以问"为什么",也可以随时停下:进度会记在工作文件夹的 进度.md 里,下次接着做。
It lays out the route, then starts at step one. Ask "why" at any step, or stop whenever you like: progress is kept in 进度.md in the work folder, so you can pick up next time.