You arrive with the app installed. In ninety minutes you'll make a webpage by asking for it, watch a vague request go wrong, understand exactly why it went wrong — and leave with a written design you approved, before a single line of code exists.
来的时候,app 要装好。这九十分钟做三件事:开口要一个网页,看它做出来; 再提一个含糊的要求,看它出岔子;然后弄明白岔子出在哪里。 散场时带走一份自己点过头的设计文档——那时候,一行代码都还没写。
Not set up yet? Do Start here first — about twenty minutes, and it's the whole course in miniature. Everything below assumes you have a paid plan, the desktop app, and a folder called my-website open in it.
还没装好?先去做 Start here,大概二十分钟。那一页是这门课的缩本。下面的内容都假设三件事已经妥了:付费账号有了,桌面 app 装了,里面开着一个叫 my-website 的文件夹。
You are in the Code tool now, not a chat — and that is the whole difference. In a chat, you type and it types back, and whatever it gives you, you copy and paste somewhere yourself. The Code tool is the same intelligence with one change: it opens the files on your computer, changes them, and saves them. You copy nothing.
我们平常说“用 AI”,其实指的是两件不太一样的事。一件是你打字,它回话;你要一段文字、一个主意、一个解释,它给你,你复制下来,再粘到要用的地方去,这是 chat。另一件是它直接动你电脑里的文件:打开、改、存,中间没有复制粘贴,这是 Code。
Not smarter. Differently placed. A chat sits on the other side of a counter and hands things over it. The Code tool is at the desk, with your work.
同一个脑子,差别不在聪明,在位置。chat 隔着一个柜台,东西从柜台递过来,进不进你的屋子,得你自己搬;Code 是进了屋的,就在你桌上做活。
Which one, when: asking a question, drafting, thinking out loud — chat. Anything where the answer has to end up in a file — Code.
这个区别决定了什么时候用哪个。问事情、起草、想主意,柜台就够了;结果最后要变成文件的,才让它进屋。
With my-website open, type this and press Enter:
打开 my-website,把下面这段打进去,回车。
Create a file called hello.html with a page that says
"Hello, my name is [YOUR NAME]" in big letters, centered,
on a dark background. Then tell me the full path to the
file so I can open it.
The app shows you exactly what it's about to change and asks permission.
app 会先告诉你它打算改什么,再问你准不准。
Say yes. Then find hello.html in your my-website folder and double-click it. It opens in your browser.
点准。到 my-website 文件夹里找到 hello.html,双击,浏览器就开了。
That's it. You just made a webpage without writing code.
就这样。一行代码没写,网页做出来了。
Remember this feeling for about four minutes. It's the last time in this course you'll ask for a finished thing and get one.
这个感觉记住四分钟。这门课里,张口要一个成品、还真拿到手,刚才是最后一次。
Same file, same agent. Ask for something everyone asks for:
同一个文件,同一个 agent。现在提一个人人都会提的要求:
Make it look nicer.
Look at what came back. It's something. It's probably even fine. But it isn't what was in your head, because nothing that was in your head ever left it. It didn't read your mind. It guessed, and it guessed like an average of the entire internet.
“让它好看点”——这句话我们平时对人说,是完全说得通的。因为对方知道这是什么场合、你是什么人、上回为什么不满意; 这些他都知道,所以他能替你补上没说出口的部分。对着这台机器说同一句话,补的人换了。 它不知道场合,不认识你,也不记得上回。它照样要补——补的办法,是拿全世界的平均水平来填。
Now the same request, said properly:
同样的要求,换个说法:
Change the background to near-black, the heading to 96px in a light grey,
and add 40px of space above and below it.
Two requests, same agent, one minute apart. The difference in what you got back is the entire course. The fix was never a better AI. It was a better instruction — and the rest of today is about where good instructions come from, because "just be more specific" is not an answer when you don't yet know what you want.
于是你拿到的东西是这样的:合理,像样,跟你没有关系。 差的不是 AI 的本事,是那句话里缺的东西。今天剩下的时间,讲这些缺的东西从哪里补—— 因为你自己都还不知道要什么的时候,“说具体一点”帮不了你。
The trap: "make it look nicer" feels like a small, reasonable request. That's exactly why it's dangerous. Nobody skips the method on a big scary project. They skip it on the one that felt easy.
陷阱在这里。“让它好看点”听着是个小要求,也很合理。可是又大又难的项目,没人会省掉方法。省掉的,永远是那个看着很容易的。
Rules 01, 02, 04 and 05 hold for good. Rule 03 holds for this course, and has an advanced version for later, which lets you do more and adds precautions to match: Reference → Advanced.
第 01、02、04、05 条一直要守。第 03 条在这门课上没有商量余地,以后有一个进阶版本:能做的更多,要做的防护也跟着更多。参考页 → 进阶。
Read the permission prompts. Don't understand one? Answer no and ask.
看清每一个授权提示。看不懂就选 no,然后问人。
Never paste a password, card number, or key into the agent. Not once, not “to test.”
绝不把密码、银行卡号或密钥贴给它。一次也不行,“试一下”也不行。
Only ever open your project folder. Never Documents, never your home folder.
只打开项目文件夹。Documents 和用户主目录不要动。
It's confident when it's wrong. Confidence isn't evidence — you check the work.
它出错的时候一样笃定。语气不是证据,得你自己去验。
If it asks you to run something from a website you didn't choose — a link it found, a script someone posted — stop and ask.
它要是让你去跑某个你没挑过的网站上的东西(它自己找来的链接,别人贴的脚本)先停下来问人。
One more, because people get it wrong in the other direction: a black rectangle drawn over a name or a number in a PDF is not a redaction. The text is usually still underneath it, and anyone can get it back. Print the page, black it out with a marker, and scan it again — or have the agent write you a small program that reads the file and produces a new one with the text actually gone.
还有一条,因为很多人是反着做错的:在 PDF 上画一个黑框盖住名字或数字,不算遮掉。文字通常还在框底下,谁都能弄回来。把那页打印出来,用马克笔涂黑,再扫一遍——或者让 agent 写一个小程序,读这个文件,生成一个真把文字去掉的新文件。
None of these are bugs, and none of them are getting fixed. You design around them.
这三条不是毛病,也不会有人去修。它们是这台机器的性质,就像水往低处流不是水的毛病。等它改是等不到的,只能绕着它安排自己的做法。
Start from what it is not. A calculator does what you tell it, and so does a dishwasher, and so does a car. Those are tools, and they almost never surprise you. This one works from vague instructions, which means everything you leave unsaid it decides for itself. Two beginner rooms, two days apart, reasoned their way to that on their own. Which is why this course treats the whole thing as management: you are not here to learn to type code, you are here to be the boss of something that types code, and you manage it the way you would manage someone you have just hired.
先说它不是什么。计算器你说什么它做什么,洗碗机、汽车也一样。那些是工具,几乎不会让你意外。这台是从模糊的指令开工的,也就是说,你没说出口的地方,它自己替你定了。相隔两天的两个新班,各自都讲到了这一层。所以这门课干脆把它当成管理来教:你来这儿不是学打代码,是学怎么当一个会打代码的东西的老板——像管一个刚招进来的人那样管它。
Put those three together and “make it look nicer” never had a chance. It forgot your last conversation, it filled every blank you left with the most likely thing rather than the thing you wanted, and it delivered that guess in exactly the same confident voice it uses when it's correct.
这三条加在一起,“让它好看点”本来就没有指望:上回聊过什么,它忘了;你留下的每一处空白,它都拿最可能的那个填上, 不是你想要的那个;递过来的是一个猜测,用的却是答对时候的语气。
Which brings you to the one rule this whole course is built on, and the one thing your instructor will be annoying about for four weeks:
这就带出这门课唯一的一条规矩。你的老师会为这件事念叨四个星期:
You don't write the documents. You cause them to exist, and you approve them. The design and the plan matter enormously — they're what keeps the agent honest between sessions. But you never type them. You talk, it writes them down, you read what it wrote and say yes, no, or “not that part.” Your four verbs: describe · decide · approve · verify. The full five moves are on the method page.
文档不是你写的,是你让它出现,然后你点头。这中间有一个分工:想清楚是你的事,写下来是它的事,认不认这份东西,又是你的事。隔一个星期还能管住 agent 的,就是这份设计和计划。四个动作:描述 · 决定 · 批准 · 验证。完整的五步在 method 页。
Say what you want. Just don't let it start building. Ask it to interview you first. In your own project folder:
我们习惯了把要求一次说完,然后等结果。这一回反过来:先别让它写,让它来问你。在自己的项目文件夹里:
I want to build a personal website for myself. Don't write any code yet.
Interview me — one question at a time — until you understand who it's for
and what it needs to do. Then propose 2 or 3 different approaches with the
trade-offs of each, and tell me which you'd pick and why.
Answer the questions. It will ask things you hadn't considered — that's the value, not a delay. When it offers approaches, pick one and say why. Then:
把问题答完。它会问到你没想过的地方。那正是它值钱的地方。等它给出几个方案,挑一个,说清楚为什么挑它。然后:
Good. Write that up as a design document called design.md: the problem,
the approach we picked, what's in scope, and what we are deliberately NOT
building. Then stop and let me read it.
Read it. Actually read it — this is the moment the whole course exists for. Then:
读一遍。是真读,不是扫一眼。然后:
Approved. Don't start building yet.
Why “what we are NOT building” earns its place: it's the line that stops the agent helpfully adding a blog, a shop, and a newsletter signup you never asked for.
“我们不做什么”这句为什么非写不可?不写,它会很热心地给你加上博客、商城、邮件订阅。这些一样都没写进去。
You are leaving today with a design and no code. That is the point. Next session, someone else's AI is going to challenge this document, and you are going to rule on every objection it raises.
今天带走的是一份设计,没有代码。这就是目的。下节课,别人的 AI 会来挑这份文档的毛病。它提的每一条,由你来裁定。
You don't have to type any of this. Both tools take dictation, and this course suits it unusually well — the method is a conversation. You describe, you decide, you approve. That's talking work, not typing work.
这些都可以用说的。两个工具都能听写。这门课特别合用,因为整套方法本来就是一场对话: 描述、决定、批准——靠嘴的活,不是靠手的活。
Claude on Mac: press Caps Lock to start dictating, again to stop.
Mac 上的 Claude:按一下 Caps Lock 开始说,再按一下停。
ChatGPT / Codex: tap the microphone icon in the text box.
ChatGPT / Codex:点输入框里的话筒图标。
Mac: System Settings → Keyboard → Dictation → on.
Mac:System Settings → Keyboard → Dictation,打开。
Windows: click a text box and press Win + H.
Windows:点一下输入框,按 Win + H。
Read it before you press Enter. Dictation mishears — say “index dot HTML” and it may write “in next dot HTML,” and the agent will do exactly the wrong thing with total confidence.
按回车之前先看一眼。听写是会听错的。你说“index dot HTML”,它可能写成“in next dot HTML”;然后 agent 会非常笃定地去做一件完全错的事。
Never dictate a password or card number. Your voice is sent away to be transcribed, not processed on your laptop.
密码和卡号绝对不要用说的。你的声音要送出去转成文字,不是在你自己电脑上处理的。
This is the week you build it. Next session puts your site on the internet, and it can only publish what exists. Unfinished is fine and expected — nobody arrives with a finished site. Arriving with nothing is the only way to waste the session.
这一周就是你动手做的那一周。下节课要把你的网站放到网上,而它只能发布已经存在的东西。没做完没关系,也很正常,谁也不会带着一个完成品来。唯一会浪费掉那节课的,是什么都没带。
Task 0, and it is not optional — do it tonight. Create your GitHub account and click the verification link in the email, then sign up for Vercel with “Continue with GitHub.” Next session is dead in the water without it: GitHub blocks repository creation and the Vercel connection until your email is verified. There is no workaround in class. Full steps →
第 0 条,不是可选的,今晚就做。注册 GitHub 账号,点开邮件里的验证链接,然后用“Continue with GitHub”注册 Vercel。少了这一步,下节课整节都动不了:邮箱没验证之前,GitHub 既不让你建仓库,也接不上 Vercel。课上没有别的办法绕过去。完整步骤 →
You have the design. You are not starting from a blank page, and you are not deciding anything new — you already made the decisions; this week is execution.
设计你已经有了。你不是从白纸开始,也不用再定什么新东西——决定你都做完了,这一周是执行。
1 · Make it write the plan. Not the code. The plan.
1 · 让它写计划。不是写代码,是写计划。
Read design.md. Write plan.md: the tasks to build it, in order, smallest
first. Each task must be one thing I can check in a browser in under a
minute. Do not write any code yet.
2 · Cut it. Read the plan and delete. Any task you couldn't check with your own eyes in a minute is too big — send it back and ask for it in halves. This is the one step people skip, and it is the one that decides whether the rest of the week goes well.
2 · 砍它。把计划读一遍,然后删。任何一条你没法在一分钟内用自己的眼睛验完,就是太大了——退回去,让它拆成两半。这一步是最多人跳过的,也是决定这一周顺不顺的那一步。
3 · One task at a time. Never “do the whole plan.”
3 · 一次一件。永远不要说「把整个计划做了」。
Do task 1 from plan.md. Only task 1. Then tell me exactly what to look at
to check it, and wait.
Then go and look with your own eyes. Never ask “did it work?” — you will get yes. Tick the task off yourself, in the file, only once you've seen it.
然后用自己的眼睛去看。永远不要问「成了吗?」——你会得到「成了」。看见了,才由你自己去文件里把那条勾掉。
4 · Two strikes, then stop. If it fails the same task twice, do not ask a third time in different words. That is the moment to change the approach, not the phrasing:
4 · 两次就停。同一条任务失败两次,别换个说法问第三遍。那一刻该换的是路子,不是措辞:
This has failed twice. Stop. Don't try again. Tell me what you think is
actually wrong, and give me two different approaches to get around it.
When do you stop building? When the page says who you are, shows what you want shown, and you would not be embarrassed for one specific person to open it. That's it. Not when the plan is fully ticked — most people arrive with a plan half ticked, and next session publishes what exists.
做到什么程度算停?这一页说清楚了你是谁、放上了你想放的东西,而且某一个具体的人打开它你不会难为情。就这样。不是等计划全部打勾——多数人来的时候计划才勾了一半,下节课发布的就是现有的这些。
Stuck all week? Come anyway, and bring the failure. A stuck project in the room gets unstuck in ten minutes; a stuck project at home stays stuck. Every prompt on this page, in one place →
整周都卡着?照样来,把卡住的地方带来。一个卡住的项目拿到屋里,十分钟就通了;留在家里,它会一直卡着。这一页上所有提示词,汇总在这里 →
Between sessions, the room is online. Each session's tips get posted to the course forum so nobody has to take notes, and questions can go there rather than waiting a week. If conversations are already piling up on you, start with where your conversations live →
两次课之间,屋子在线上。每节课的要点都会发到课程论坛,谁都不用记笔记;问题也可以直接发在那里,不必等一个星期。要是对话已经堆得你头疼,先看你的对话放在哪里 →
Close the app completely. Open it again, open my-website, and ask, cold: “Read design.md. What are we building?” It should answer correctly with nothing from you. That's the whole reason the document exists.
把 app 完全关掉。重新打开,打开 my-website,劈头问一句:“Read design.md. What are we building?”你不补一句背景,它也应该答得对。这份文档存在,就是为了这一问。
Read design.md once more, a day later. Mark one sentence you're proud of and one sentence that's vague. You'll spot the vague one much more easily now than you could five minutes after writing it.
过一天,把 design.md 再读一遍。标出一句写得好的,和一句写得含糊的。含糊的那句,隔一天去找,比刚写完五分钟容易得多。
Pick the part of your design you're least sure about, and write down one sentence about why. Keep that sentence — in session 3 you'll hand it to a classmate's AI and ask it to be blunt. Write it down now, while you still remember which part worried you.
挑出设计里你最没把握的一处,写一句话说明为什么没把握。这句话留着——第三次课交给同学的 AI,让它说狠话。现在就写下来,趁你还记得是哪一处让你不放心。
3–5 sentences about yourself · 1–2 photos you're happy to have public · your email address as you want it shown. You're building the real thing this week, and it goes faster if the words already exist.
三到五句自我介绍 · 一两张你愿意公开的照片 · 你希望显示出来的邮箱。这一周就要做真东西,字先写好,就快得多。
If the cold-start test fails — it answers wrongly, or asks you what you're building — that's not a problem, it's a finding. Bring it. A design that can't survive a closed window is exactly what session 2 is about fixing.
要是这个冷启动测试没过(它答错了,或者反过来问你要做什么)那不算出问题,那是查出了一件事。带来。一份关掉窗口就撑不住的设计,正好是第二次课要修的东西。
Session 2: whatever you've built by then goes online, at an address anyone on earth can open.
第二次课:到那时候你做出了什么,就把什么放上网,一个地球上任何人都能打开的地址。