Session 4
第四次课

Your own idea

你自己的想法

Three sessions ran the loop on a website. Today you bring something you actually want, and run the loop up to the point where building would start. Nothing today depends on writing code — start task one at the end if you have time, but you leave with a plan somebody else has already pushed back on, and with your own method written down.

前三次课是拿一个网站把这套循环跑了一遍。今天你带一个自己真想做的东西来——把循环走到「该动手做了」为止。今天没有哪一步是靠写代码的——最后有时间可以把第一条做了,但你带走的是一份已经被别人推敲过的计划,以及你自己那套方法,写下来的。

The takeaway is the method, not the thing. Whatever you built in the first three sessions goes out of date, and so will whatever you build next. A way of working doesn't expire the same way — but it does need maintaining, which is the second question of the day and the one people never ask themselves.

你带走的是方法,不是那个东西。前三次课做出来的东西会过时,接下来做的也会。一套干活的方式不会那样过期——但它需要维护,而这正是今天的第二个问题,也是大家从来不问自己的那个。

So there are two questions, not one: do you have a way of working, and can you say it out loud? Then — how do you keep it current when the work gets harder and the models get better? The second one is why this isn't just a show-and-tell.

所以是两个问题,不是一个:你有没有一套自己的干活方式,而且你能不能把它说出口?然后——等活变难了、模型变强了,你怎么让它跟得上?第二个问题,是这一次课不只是「各自展示一下」的原因。

Before you walk in
进门之前

Bring one idea, written down

带一个想法来,写下来

Not a category — an idea. Not "something for my business." A one-page menu for the supper club I run, that I can change from my phone. Three or four sentences is plenty. If you can't write it down, you can't hand it to an agent either, and you'll spend the first twenty minutes discovering that in public.

不是一个门类,是一个想法。不是“给我生意做点东西”,而是我办的那个家宴,做一页可以从手机上改的菜单。三四句话就够。写不下来,你也就交不给 agent;那你会在头二十分钟里当众发现这件事。

It does not have to be a website, and it does not have to involve AI. An app. A newsletter. A business you've been circling for two years. A short film. The loop you've been running is not a website loop — describe it, get it written down, have it challenged, break it into pieces you can check. Run it twice on different things and you find the same shape underneath both.

不一定非得是网站,也不一定跟 AI 有关。一个 app。一份通讯。一个你想了两年的生意点子。一部短片。你一直在跑的这套循环不是「网站专用」的——把它说清楚、让它被写下来、让它被挑毛病、切成你验得了的块。换两件不同的事各跑一遍,你会在底下看见同一个形状。

Which also means the constraint from the last three sessions is lifted. Logins, payments, databases, other people's services — today you can plan any of them, because planning something is not the same as having ninety minutes to debug it. What you cannot bring is a category. "Something for my business" is not an idea; the supper-club menu is.

这也意味着前三次课那条限制解除了。登录、支付、数据库、别人家的服务——今天这些你都可以拿来做计划,因为「给一件事做计划」和「有九十分钟去给它排错」不是一回事。不能带来的是一个门类。「给我生意做点东西」不是一个想法,那份家宴菜单才是。

Three real ones, for scale. A browser extension that drafts replies to posts in a few set styles, so a person only has to pick one. A heap of company files scattered across a laptop and a shared drive, turned into a set of notes with an index over them. A meeting recording that comes back as a list of tasks, which somebody then checks against what was actually agreed. Not one of them is a website. Each one is written up in the course forum, with how it was actually built.

三个真的例子,给你一个尺度。一个浏览器插件,按几种预设风格给帖子生成回复草稿,人只需要挑一条。一堆公司文件,散在笔记本和共享盘上,整理成一组笔记加一份索引。一段会议录音,回来变成一张任务清单,再由人对着会上真正说定的内容核一遍。没有一个是网站。每一个在课程论坛里都写下来了,连怎么做的一起。

If your idea is a person, know which of two projects you are asking for. An AI of a grandmother; a tutor that teaches the way a particular professor teaches. These arrive more often than any other kind and they are usually the ones their owners care about most, which is exactly why it is worth an hour, before you walk in, to work out which of the two you actually want. They cost different amounts by a very wide margin.

如果你的想法是「一个人」,进门之前先弄清你要的是两件事里的哪一件。一个我奶奶的 AI;一个像我某位老师那样教书的助教。这类想法来得比任何一类都勤,而且往往是提的人最上心的那个,正因如此,值得在进门之前花一个钟头想清楚你要的到底是哪一件。这两件事的价钱差得很远。

Taking a general model, giving it what the person actually wrote, and asking it to answer in their manner is something you can try this week for the price of a subscription. Training a model that is that person is a hardware bill followed by a hosting bill, at a scale people raise money for. Nearly every idea of this kind wants the first and asks for the second, and the asking is usually done in good faith, by someone who has no reason yet to know there are two things to ask for.

拿一个通用模型,把这个人真正写过的东西喂给它,让它按这个人的路数回答,这件事你这个星期就能试,成本是一份订阅。从头训一个「就是这个人」的模型,那是一笔硬件账后面跟一笔托管账,是要去拉投资的数量级。这类想法几乎都是想要前者、开口要后者,而且开口的人多半是诚心的,只是还没有理由知道这里头是两件事。

So come with two answers ready. The first is what material of theirs you actually hold: not what exists somewhere in the world, but what sits on a disk you can open tonight. Their published work is the easy half, and it is rarely the useful one; the useful half is the unglamorous pile — the feedback, the replies, the drafts they marked up for other people — because a manner of working is visible there, being practised, rather than described. The second answer is whether you have asked them. Someone in the room will raise it, and the question lands more gently coming from you.

所以带两个答案来。第一个是这个人的材料你手上真正有多少:不是世上存在多少,是今晚就能打开的硬盘上有多少。他发表的东西是容易的那一半,而且往往不是有用的那一半;有用的是不起眼的那一堆,批注、回信、他替别人改过的稿子,因为一个人做事的路数是在那里被实践出来的,不是被描述出来的。第二个答案是你问过本人没有。屋里总会有人提起这一句,由你自己先说出来要好受些。

And write down how you work

还有,把你的干活方式写下来

The second half of the homework, and the one people skip. In your own words, on one page: given a thing you want to exist, how do you use AI to get it made? Not the version from these pages — yours, including the steps you dropped and the ones you added because they kept saving you.

作业的另一半,也是大家会跳过的那一半。用你自己的话,写一页:给你一件你希望它存在的东西,你怎么用 AI 把它做出来?不是这些页面上的那一版——是你的,包括你砍掉的步骤,和你自己加上、因为它一直在救你的那些。

If yours comes out very short, that is a good sign, not a weak one. A method you actually follow beats a thorough one you don't. Four lines you use every time is a real answer; the honest test is whether you could hand it to someone and have them work the way you work.

写出来要是很短,那是好事,不是单薄。一套你真会照着做的方法,胜过一套详尽但你不用的。四行、每次都用得上,就是一个真答案;诚实的检验标准是:把它交给别人,对方能不能照着你的方式干活。

Then add two lines at the bottom, because a method that never changes is a method you've stopped looking at:

然后在最下面加两行,因为一套从来不变的方法,等于你已经不再看它了:

WHAT CHANGED
变了什么

One thing you do differently now than in session 1, and what made you change it. If nothing changed, say that — it's an answer, and usually an uncomfortable one.

有哪一件事,你现在的做法和第一次课时不一样了,以及是什么让你改的。要是什么都没变,就写「没变」——那也是一个答案,而且通常是个让人不太舒服的答案。

WHAT YOU'D REVISIT
什么时候该回头看

Which step you'd expect to drop first if the models got noticeably better. Naming it now is what stops you still doing it in two years out of habit.

要是模型明显变强了,你觉得哪一步会最先被你砍掉。现在就把它点出来,你才不会两年后还因为习惯在做它。

How the session runs
这次课怎么走

You pitch it, the room takes it apart

你讲,屋里的人拆

Take the people in this room as the people you'd need money from. You get a few minutes to say what you want to make, how you'd go about it, roughly how long, and roughly what it looks like when it's done. Then it gets challenged — same as session 3, except this time nothing has been built yet, so what's on trial is the plan.

把在座的人当成你要去拿钱的人。你有几分钟,讲清楚:你想做什么、你打算怎么做、大概多久做出来、做出来大概是什么样子。然后它就会被挑毛病——和第三次课一样,只不过这一次什么都还没做出来,受审的是那份计划。

Your pitch is Move 1, out loud. Same four things the interview drags out of you — who it's for, what it does, what's in scope, what you're deliberately not doing — so once you've said it to the room and survived the questions, you hand it to your agent and have it written up as design.md. That's the file the rest of the session runs on. If you can't get through the pitch, you were never going to get through the interview either, and finding that out in four minutes is the cheapest version of that news.

你的讲述就是第一步,只不过是说出来的。问询要从你嘴里挖的也是那四样——给谁用、它干什么、做什么、有意不做什么——所以你对着屋里讲完、扛住提问之后,把它交给你的 agent,让它整理成 design.md。这一次课后面的部分,就跑在这个文件上。要是你讲都讲不下来,那你本来也过不了问询那一关,而用四分钟发现这件事,是这个消息最便宜的版本。

1 pitch it 讲你的想法 what, how, how long 做什么、怎么做、多久 2 challenged 被挑毛病 the room and its AIs 屋里的人和他们的 AI 3 compare methods 比方法 yours, theirs, mine 你的、他们的、我的 4 plan.md you refuse the big ones 太大的你退回 the next step 下一步 in writing 写下来 CODE IS OPTIONAL 写不写代码随意
Stops 1, 2, 4 and 5 you have run before. Stop 3 is the one that only works with other people.
第 1、2、4、5 站你都跑过。第 3 站是那种只有身边有别人才成立的。

These are stops, not a timetable. People arrive with genuinely different amounts done, so nobody is pacing the room to a clock. None of the five depends on writing code. Building is what you do after class; the thing that goes wrong when you do it alone is starting badly, and that is the part worth spending a session on.

这是几站,不是时间表。每个人到场时手里的进度差得很远,所以没有人拿着秒表带着全屋走。这五站没有一站是靠写代码的。做,是下课之后的事;一个人做的时候真正会出问题的是开头开坏了,而那一段才值得花一次课。

When it's someone else's turn
轮到别人讲的时候

Challenge the idea without rebuilding it

挑这个想法,但别替对方重做一个

This goes wrong in one specific way and it's worth naming first. In session 3 the failure was the confident person taking the keyboard. Today there's no keyboard, so it arrives as: the confident person explains what you should be building instead. Five minutes later the room is discussing their idea, and the person whose turn it was has stopped talking.

这件事会以一种特定的方式走坏,值得先点出来。第三次课那个走法是「有底气的人把键盘接过去」。今天没有键盘,于是它变成:有底气的人开始讲你应该做的是别的什么。五分钟后,屋里在聊的是那个人自己的想法,而本该轮到的那个人已经不说话了。

The same three questions do the work, and they are all about their idea:

还是那三个问题最管用,而且都是围着对方的想法问的:

WHO IS IT FOR
给谁用

Name one real person who would use this. If nobody can, that's the finding — and it's better heard now than after a month of building.

说出一个真会用它的人。说不出来,那就是那条意见——现在听见,好过做了一个月之后。

WHY THIS SHAPE
为什么是这个形式

Website, app, newsletter, video — why that one? This is wrong-medium asked before it's expensive instead of after.

网站、app、通讯、视频——为什么是这一个?这就是把 wrong-medium 提前到还不贵的时候问,而不是等到之后。

WHAT'S FIRST
第一刀切哪

What's the smallest piece that would be worth having on its own? A plan whose first step is “build it” isn't a plan yet.

最小的那一块、单独拿出来也值得有的,是哪一块?第一步写着「把它做出来」的计划,还不算计划。

And run them past your own AI too, with a person assigned, exactly as in session 3 — an investor, a first customer, someone who already sells something like this. The owner rules on every finding. Not you, and not your model.

同样也用你自己的 AI 跑一遍,先指定身份,和第三次课一模一样——一个投资人、一个最早的客户、一个已经在卖类似东西的人。每一条意见由想法的主人来裁。不是你,也不是你的模型。

Then the part that has no equivalent anywhere else: everyone's written method goes side by side. Yours, the other people's, and your instructor's, plus at least one from outside the room — the team that made a feature-length film with AI published theirs, as an example of a production method written down. Read several and you can see which parts of your own are habit and which are load-bearing. You cannot get that from one.

然后是别处没有的那一段:把每个人写下来的方法并排放在一起。你的、别人的、你老师的,再加上至少一份屋外的——那支用 AI 拍完一部长片的团队把他们的方法公开了,那就是一份写下来的方法论的样子。多读几份,你就能看出自己那份里哪些是习惯、哪些是承重的。只读一份,看不出来。

The last thing that isn't optional
最后一件非做不可的

Turn what survived into a plan

把活下来的东西变成计划

Your pitch was the interview, spoken out loud, and the challenge round already made you rule on what came back. That is design.md in everything but the file. Write it down, then get the plan — the same two prompts as every other session, because it is the same loop.

你的讲述就是问询,只不过是说出来的;被挑那一轮,你已经对拿回来的意见一条条裁过了。那就是 design.md,只差没落成文件。把它写下来,然后拿计划——和别的每一次课是同样两段提示词,因为是同一个循环。

Decide which of two jobs you are doing, before you start. Take its idea and let it finish, and a few rounds will do it. Hold a precise picture in your head and insist on that exact thing, and expect roughly 80% of your time to go on closing the last 20% of the gap. Both are legitimate choices. The second one needs the picture written down, because it cannot see the one in your head, and it needs the time budgeted, because that is where the time goes.

开工之前先定:你做的是两件事里的哪一件。拿它的主意,顺着它做完,来回几轮就够了。你脑子里已经有一个很具体的样子,非要做成那个样子,那就要准备好大约 80% 的时间花在最后 20% 的差距上。两种都算正当选择。第二种要把那个样子写下来,因为它看不见你脑子里的;也要把时间留出来,因为时间就花在那里。

Write that up as design.md: the problem, the approach we picked, what's in
scope, and what we are deliberately NOT building. Add a "Second opinion"
section with each finding and what I decided: accepted, dropped, dissolved,
or wrong-medium, with my reason. Then stop and let me read it.
Turn the approved design.md into an implementation plan in plan.md.
Break it into small tasks. For each task: what changes, the exact steps, and
how I check with my own eyes that it worked. Number them. Then stop.

Send at least one task back. Go down the list and ask one question of each: could I check this with my own eyes in about a minute? If you think the plan was perfect, you didn't read it closely — that is the single most reliable thing anyone has learnt about plans in this course. The full version of this step →

至少退回一条任务。一条一条往下过,每条只问一句:这一条我能用自己的眼睛、在一分钟里验完吗?你要是觉得这份计划完美无缺,那是你没细读——这是这门课关于计划最靠得住的一条经验。这一步的完整版 →

Optional · only if the planning is genuinely done
选做 · 只有计划真做完了才轮到它

If there's time left, start task one

要是还有时间,就把第一条做了

Some people arrive with a plan already sharp and finish the pitch and the challenge with half an hour to spare. If that's you, build task one — and only task one — the same way you did in the earlier sessions: it does the task, then stops and tells you exactly what to open, and you check with your own eyes before anything else happens.

有的人来的时候计划已经很利落了,讲完、被挑完,还剩半小时。要是你是这种,就把第一条任务做掉——而且只做第一条,和前几次课一样:它做完停下来,告诉你打开哪里看,你先用自己的眼睛验过,再说别的。

Do task 1 from plan.md. Only task 1. Then tell me exactly what to look at
to check it, and wait.

Optional means optional, and the order is not negotiable. Nobody is behind for skipping this. Building is the part you can do alone on a Saturday; the pitch, the challenge and the method comparison need a room, and a room is the scarce thing. If you're choosing, choose the parts you can't do at home.

选做就是选做,但顺序不能换。跳过这一段的人不算落后。做,是你周六一个人也能做的;讲、被挑、比方法,得有一屋子人,而一屋子人才是稀缺的那样东西。要取舍的话,先要那些你在家做不了的。

And if it fails twice, stop — the two-strike rule holds here exactly as it did in session 3. A third attempt at the end of a session, with everyone packing up, is the worst possible moment to be making a mess you'll inherit next week.

要是失败两次,就停——「两次就停」这条规矩在这儿和第三次课完全一样。一次课快散场、大家都在收东西的时候试第三遍,是给下周的自己留烂摊子最糟糕的时机。

The finish line
终点线

What “started” actually means

“开起来了”到底是什么意思

“I started a project” is the easiest sentence in the world to say and mean nothing by. Here it means four things exist, and you can point at all of them:

“我开了一个项目”是世界上最容易说、又最可能什么都不代表的一句话。在这里,它的意思是有四样东西存在,而且你能一一指出来:

01

An approved design.md

一份批准过的 design.md

With a “Second opinion” section in it, showing what the room and its AIs found, and what you decided about each finding — accepted, dropped, dissolved, or wrong-medium.

里面有一节叫“Second opinion”,写着屋里的人和他们的 AI 找出了什么、你对每一条的裁定是什么——accepted、dropped、dissolved 还是 wrong-medium。

02

A plan.md you cut down

一份被你砍过的 plan.md

Numbered tasks, each small enough to check by eye. At least one task you sent back for being too big.

编号的任务,每一条都小到能用眼睛验。其中至少有一条是你因为太大而退回去的。

03

Your method, in your own words

你自己那套方法,你自己的话

One page: given a thing you want to exist, how you use AI to get it made. Revised after hearing other people's.

一页纸:给你一件你希望它存在的东西,你怎么用 AI 把它做出来。听过别人的之后,改过一遍。

04

The next task, in writing

下一条任务,写下来

One sentence at the bottom of plan.md saying what you do first when you sit down alone on Saturday.

在 plan.md 最下面写一句话:周六一个人坐下来的时候,第一件事做什么。

Number four is the one that matters after today. The reason people stop is not that it got hard — it's that they sat down, couldn't remember where they were, and did something else instead. A written next task is the difference between a project and a nice afternoon.

第四条是今天过后真正起作用的那一条。人们停下来,不是因为变难了——是因为坐下来之后想不起自己到了哪儿,于是去干别的了。一条写下来的下一步,是「一个项目」和「一个愉快的下午」之间的区别。

After the course
课程之后

What you actually walk out with

你真正带走的东西

A live website with your name on it, and a second project started properly. But the thing worth keeping is smaller and more portable than either:

一个挂着你名字的线上网站,加一个正经开起来的第二个项目。但真正值得留下的东西比这两样都小、也都更好搬:

You know how to be the person who decides. Describe it. Get it written down. Have someone else challenge it. Approve it. Break it into pieces you can check. Look with your own eyes. Stop after two failures and re-plan.

你知道怎么当那个做决定的人。描述它。让它被写下来。找别人来挑它。批准它。切成你验得了的块。用自己的眼睛看。失败两次就停下来重做计划。

None of that is about websites, and none of it expires when the tools change. The tools will change — probably before you finish the project you started today. The loop won't.

这些都不是关于网站的,也不会因为工具换代就过期。工具是会换的——很可能在你把今天开的这个项目做完之前就换了。循环不会。