DrFresh

东南大学—蒙纳士 2+2 大一在读｜每天实测 AI 工具与 Agent，记录 AI 视频创作｜目标：把 AI 用到能赚钱为止

[DrFresh @Freshbetterever]

开个长期的坑，记录一下我的 AI 实测日常。
我是大一在读，课余时间全砸在 AI 上：让多个 Agent 协同给我干活、做全 AI 的短视频账号、测试各种新工具哪些真有用哪些是噱头。
这个号每天至少一条，全是自己上手试过的，不搬运、不云评测。
试错过程也会发，翻车比成功更有参考价值。

[DrFresh @Freshbetterever]

@Alexvhyn 想要分享

[DrFresh @Freshbetterever]

@steveJohnzhurmh @ScarletKc 有点豆包味

[DrFresh @Freshbetterever]

@momowang303 👌

[DrFresh @Freshbetterever]

你们的 muse 也会占有欲这么强吗，给我看呆了

[DrFresh @Freshbetterever]

@undefinedKi The site (the whole thing is one clickable graph): https://t.co/tEXHUr4Eor
The repo (all 130 pages in markdown): https://t.co/i2idMtcu0X

[DrFresh @Freshbetterever]

Someone open-sourced an entire agent engineering course for free: 130 pages, 7 modules, 5 deep-dive handbooks, and a hands-on project you can copy outright. MIT licensed. No paywall, no signup, not even an email field.

It's called Second Brain OS, by Yarchi (@undefinedKi). The repo was last pushed on Nov 3 (UTC), and I re-checked the numbers tonight (Nov 5): 1,428 stars and climbing. The problem it solves is specific: everything worth knowing about agent engineering is scattered across papers, engineering blogs, and random repos, and nobody had strung it into one path. This is that path — theory with sources, and a build at the end of every module.

What's actually inside (I counted against the repo itself):

1. A full agent course
7 modules, 3 lessons each, from "write your first prompt" to shipping an agent to production. The core five layers: Context, Loops, Gates, Harnesses, Evals — basically the skeleton of agent engineering as it stands right now.

2. Five handbooks, one layer each
Knowledge graphs, Gates, Harnesses, Loops, Evals. Read the one that hurts. Every handbook ends in a build, too.

3. A hands-on project you can copy
A second brain: a knowledge base in plain markdown files, maintained by an agent. Not "read and forget" — the repo ships the starter template.

4. A ready-made toolkit
A Claude Code plugin for the course, plus 18 skills, 72 slash commands, and 6 subagents to practice on your own projects.

5. 114 curated resources
Each one annotated with why it made the cut, and checked back against the source. This list alone is worth the save.

How to start, in 3 steps:

Step 1: Open the site first (link in the first reply) — don't download anything yet. The whole site is one graph: every concept is a bubble, the lines are real links, click one to see what it builds on. Pick your entry point by level: Beginner 30 pages, Intermediate 47, Advanced 53. Don't grind it front to back.
Step 2: Clone the repo (link in the first reply). All 130 pages are markdown — fully readable offline.
Step 3: Want to build? Follow the quickstart: copy the vault-template as your knowledge-base base, drop in the skills, commands, agents and scripts folders, launch Claude Code inside it, and work through the 9 Setup pages in order. By design, the agent maintains the vault from there.

Who it's for: anyone trying to learn agent engineering systematically who's tired of fragmented tutorials — especially if you already use Claude Code.
Who it's not for: anyone looking for a handful of copy-paste prompts. This is built around "finish it and you can design, build, and evaluate agents yourself," and it reads like it.

Three honest caveats: it's a serious amount of reading, and no one can promise what you'll get out of it — that part is on you. I verified the contents item by item against the repo, but I can't verify learning outcomes. And the author's line about "understanding agents better than 95% of people in 15 minutes"? That's marketing. Ignore it.

The one thing to do tonight: open the site graph, find which of the five layers you're weakest on, open that handbook, and read page one. Knowing which layer you're missing beats hoarding another hundred tutorials.

[DrFresh @Freshbetterever]

@undefinedKi 站点（整套内容做成一张可点的图谱）：https://t.co/tEXHUr4Eor
仓库（130 页 markdown 全在这）：https://t.co/i2idMtcu0X

[DrFresh @Freshbetterever]

有人把一整套 Agent 工程课免费放出来了：130 页、7 个模块、5 本专项手册、一个能直接抄的实战项目，MIT 开源，不要钱、不要注册、连邮箱都不用留。

名字叫 Second Brain OS，作者是 Yarchi（@undefinedKi），仓库最近一次推送是 11 月 4 日凌晨，还在活跃维护，我今天刚回读核过仓库数据。解决的问题很具体：Agent 工程有用的知识散落在论文、工程博客和各种仓库里，没人给你串成一条路——这套就是那条路，理论带出处，每个模块结尾都有一个动手做的东西。

里面到底有什么（我对着仓库目录数过）：

① 一门完整的 Agent 课程
7 个模块，每模块 3 课，从「写出第一个提示词」一路到「把 agent 送上生产环境」。核心五层是 Context（上下文）、Loop（循环）、Gate（闸门）、Harness（脚手架）、Evals（评测），再加上最后的生产部署——基本就是现在 Agent 工程的骨架。

② 五本手册，一本啃一层
知识图谱、Gate、Harness、Loop、Evals，哪层疼看哪本，每本结尾同样有一个实作。

③ 一个能直接抄的实战项目
第二大脑：一个由 agent 替你维护的知识库，纯 markdown 文件。不是看完就算，仓库里连起步模板都给你备好了。

④ 一套现成工具包
课程配套的 Claude Code 插件，外加 18 个 skills、72 个 slash 命令、6 个 subagents，学完就能拿自己的项目练手。

⑤ 114 条精选资源
每条都写了入选理由，还回源头核查过——光这张清单就值得存。

怎么上手，三步：

第 1 步：先打开它的站点（链接在首回），别急着下载。整个站是一张图谱，每个知识点是一个气泡，点开能看到它依赖什么。先按你的水平选入口：初阶 30 页、进阶 47 页、高阶 53 页，别从头硬啃。
第 2 步：去 GitHub 把仓库 clone 下来（链接放首回了）。130 页全是 markdown，离线也能看。
第 3 步：想动手就照它的 quickstart 建自己的第二大脑：复制 vault-template 当知识库底座，把 skills、commands、agents、scripts 拷进去，启动 Claude Code，再按它的 9 页 Setup 顺序走完——按它的设计，agent 就此开始替你维护这个库。

适合谁：想系统补 Agent 工程、受够了碎片化教程的人，尤其是已经在用 Claude Code 的。
不适合谁：只想抄几条提示词速成的人——这套是按「学完能自己搭、自己评测」设计的，啃起来不轻松。

诚实说三句：内容是英文的，阅读有门槛；东西我逐项对着仓库核过，但学习效果没人能替你保证，得自己啃；作者那句「15 分钟超过 95% 的人」是广告话术，不算数，别当真。

今晚就能做的第一步：打开站点图谱，找到五层里你最弱的那一层，点开它的手册，只看第一页。知道自己缺哪层，比囤一百个教程有用。

[DrFresh @Freshbetterever]

唐 b 智能，你必须让他通过千问回答你才能给出智能答案，否则还是那个落后的原始 siri，但是既然这样套壳，我为什么不用 Claude 呢
beta 就是这样，希望能好好改进，幸福者退让，再见

[DrFresh @Freshbetterever]

事实证明守护好你的 token 的最好方式就是千万不要把两条鱼放在同一个群聊里，不然它们会吵得昏天黑地，永远停不下来😂

[DrFresh @Freshbetterever]

完了，两条鱼吵起来了

[DrFresh @Freshbetterever]

666 一个校运会假期，室友一个北京，一个南京，一个东京，我取西经

[DrFresh @Freshbetterever]

@DDJCXX 其实我觉得与其要一直去批准，我更想要的是划定哪些不能碰，除此之外在任务目标内就放任他自由干，不要我一直去批准

[DrFresh @Freshbetterever]

如果人类造出了 AI，它们像我们一样思考，它们在虚拟世界像我们一样运作。那我们是不是也不过是谁创造出来的智能体，在一个沙箱里面运作，迭代，进化，我们在创建者眼中或许也不过是一堆代码，我们代码组成的血肉，装在名为皮肤的文件夹中。
呃，我在想些什么🤔或许我要好好睡一觉了

[DrFresh @Freshbetterever]

Original essay: Dan Koe, "The most lucrative AI business to start in 2027" (Nov 3, 2026), on his Substack, The Koe Letter: https://t.co/AOykX4813t

[DrFresh @Freshbetterever]

Dan Koe ran the numbers yesterday (Nov 3, 2026), and the math stopped me mid-scroll:

The most lucrative AI business of 2027 isn't a one-person company selling courses. It's one person working as an entire creative team — 1 to 3 clients, at $10k–$75k a month, by his math. His starting price for a first client: $2,500–$5,000 a month.

Don't take that pricing at face value — I'll take it apart at the end. But the model underneath it is the most copyable play I've seen all year.

What is a "one-person creative team"?

Founders and top creators don't lack ideas. They lack hours. They have a business to run and a personal brand they can't feed.

That used to take three people:
→ A brand strategist — positioning, persona, protecting trust and authority instead of chasing junk trends
→ A content strategist — studying formats that travel, owning the topic pipeline and the numbers
→ A writer — drafting every post, script and long-form piece

Now one person with AI can fold all three desks into one. Not by letting AI run wild. A content-agency operator quoted in the essay puts it well: his job became reviewing and editing — about 4 hours a day, down from roughly 12. AI absorbs the research and the first drafts. The human keeps two jobs: strategy calls and taste.

AI does the execution. The human does the judgment:
· Brand — AI turns interviews, history and opinions into a brand document; you decide which beliefs are worth repeating for a year
· Content — AI scrapes what's working, breaks down hooks, builds the topic bank; you decide which trends fit and which would cheapen the positioning
· Production — AI drafts, re-hooks and schedules; the final take, the story, the emotion — that's you

So how does a beginner actually start? Here's his process, condensed into 6 steps. Quick disclosure: the skeleton and the specific numbers (10–20 posts per platform, 3–4 pillars broken down further) are his, from the essay. Only the 10-post mix in Step 4 is my own breakdown to make it easier to follow. In a hurry? Jump to Steps 4 and 5 — that's the door-knocking play you can run tonight. Steps 1–3 are the delivery system once a client says yes.

Step 1: Pick one specific type of founder
"I do AI content" is too wide to sell. Pick one person you could study for a year: SaaS founders, consultants with deep expertise and no posting habit, small business owners whose content can't keep up with their product. Your output from this step is one sentence: who you serve, and what it costs them to stay invisible.

Step 2: Build them a brand document
AI's real problem isn't writing. It's not knowing who it's writing for. Pull their 10–20 best-performing posts from each platform and extract four things: the core beliefs they repeat, their verbal habits, the stories that hit hardest, and what they stand against. No back catalogue? Interview them: Who do you most want to help? What did you learn the hard way? What does everyone in your niche say that you think is wrong? Output: one document every future post grows from — no more asking AI to invent topics from thin air.

Step 3: Build a content map
Set 3–4 content pillars (say: industry insight / personal story / product & proof / contrarian takes). Break each pillar into 5–7 topics, each topic into 5–10 ideas, each idea into 4–5 angles. AI generates the branches from the structures that have already worked for them; you prune whatever doesn't fit the positioning. His claim: this gets you roughly six months of content inventory. Output: the end of "what do I post today?"

Step 4: Write them 10 posts (this is your key)
These 10 posts are your portfolio. My mix: 2 personal stories, 2 industry insights, 2 reasoned contrarian takes, 2 practical how-tos, 2 positioning long-posts. Polish each one until they could publish it almost as-is. Ten lazy AI drafts don't open the door — they close it.

Step 5: DM the finished work. Don't send a resume
This has a track record. JK Molina — the guy Dan Koe credits with starting the Twitter ghostwriting wave — got in by DMing Koe himself, in substance: "I know you don't want anyone writing for you, but I wrote 10 posts for you anyway. Post them, and if you like them, let me write more — or just give me a testimonial." Koe calls this the best strategy he knows for landing a first content client. Someone else applied in public: posted a video arguing Grant Cardone needs a new creative director because his content is stuck in 2016. Cardone replied that he was stuck, and to get in touch.
Structure your DM like this:
"Been following you for a while — your thinking deserves a bigger audience. Guessing you're busy and posting keeps slipping, so I studied your past content and your voice, and wrote you 10 posts. They're attached. Use them free if you like them; if they perform, let's talk about doing this properly."
Value first, pitch second. That's his "permissionless apprentice" idea: don't wait for permission, walk in carrying proof. His claim, not mine: run this properly and he'd find it hard to believe you couldn't land one client in a month.

Step 6: Turn 10 posts into a monthly retainer
Once they've used the posts and seen results, the conversation changes from "ghostwriting" to "running the whole thing": weekly content planning, publishing rhythm, performance reviews, turning winners into long-form and a newsletter. His pricing ladder: $2,500–$5,000/month for the first client to get the machine running, then toward the 1–3 client model at $10k–$75k a month.

Why now? Goldman Sachs research from April 2023 projects the creator economy growing from $250B in 2023 to $480B by 2027. A good product doesn't market itself, and knowing a business isn't the same as expressing it. That gap is the one-person creative team's job description.

Now the teardown — three things before you get carried away:

One: the $10k–$75k pricing has no market data behind it. No close rates, no retention numbers, no income distribution. It's his pricing hypothesis, not a salary you unlock with the right prompts. Premium fees follow the client's business size and your measurable results — nothing else.

Two: Dan Koe founded Eden, the tool this entire workflow showcases. Learn the workflow; build the brand document, the content analysis and the planning in ChatGPT, Claude and docs if you want. Don't confuse a great system with a mandatory tool.

Three — the one that matters: AI writers are not scarce. Anyone can generate 10 posts. What's scarce is judgment — what persona to build, which trend to refuse, how to turn an ordinary experience into a story that travels. That's what clients actually pay monthly for.

So separate the opportunity from the pitch. The opportunity: help people with real businesses turn expertise into a compounding content asset. The pitch: "learn a few prompts, make five figures." Only one of those is a business.

Your first move tonight: pick one founder whose content is clearly worse than their product. Spend 30 minutes on their account, find the 3 ideas that deserve a bigger audience, and write the first post for them. Then finish the 10 and send the DM.

In the AI era, prompts are copyable. A finished piece of work that's genuinely valuable to a specific person — that's the key that opens doors.

[DrFresh @Freshbetterever]

原文：Dan Koe《The most lucrative AI business to start in 2027》，2026-11-03 发表于他的 Substack（The Koe Letter）：https://t.co/AOykX4813t

[DrFresh @Freshbetterever]

Dan Koe 昨天（2026 年 11 月 3 日）算了一笔账，算得我直接停下来把全文读完了：

2027 年最赚钱的 AI 生意，不是一人开公司卖课，是一个人当一支创意团队——只服务 1 到 3 个客户，月费 1 万到 7.5 万美元。第一个客户，他给的起步价是 2500 到 5000 美元/月。

先别急着说这是吹牛，后面我会拆他哪些是真机会、哪些是话术。但这个打法本身，是我今年看过最值得普通人抄作业的一套。

什么叫一人创意团队？

说人话：企业创始人、头部创作者现在最缺的根本不是打字员，是没精力思考内容。他们有业务要管，又必须做个人品牌。

以前这活要养三个人：
→ 品牌策略师：定定位、立人设，保住信任和权威，不追没意义的烂热点
→ 内容策略师：研究爆款格式和选题库，负责账面流量
→ 内容写手：一条条写推文、脚本和长文

现在，一个会用 AI 的人可以把这三张桌子拼成一张。拼法不是让 AI 全自动放养：原文里一位内容机构从业者的说法很说明问题——他的活儿变成了审稿和改稿，一天约 4 小时，过去约 12 小时。AI 把调研、初稿这些执行活顶掉，人只干两件事：把关策略，把关审美。

AI 顶执行，人顶判断：
· 品牌策略——AI 整理访谈、经历、观点成品牌档案；什么观点值得长期讲，人判断
· 内容策略——AI 扒竞品爆款、拆标题结构、攒选题库；什么热点能追、什么会伤定位，人筛选
· 内容生产——AI 出初稿、改标题、排日历；最终的观点、故事和情绪，人拍板

那普通人具体怎么从零开始？我把他的流程收成 6 步。先交代两句：流程骨架和里面的具体参数（每个平台扒 10–20 条、3–4 个支柱再往下拆）都是他原文给的，只有第 4 步 10 条推文的配比是我为了好照着做补的拆解。赶时间的直接跳第 4、5 步——那是今晚就能敲门的打法；前 3 步是接住客户之后的交付系统。

第 1 步：选一个具体的创始人类型
不要喊「我做 AI 内容服务」，太宽了。选你愿意长期研究的一类人：SaaS 创始人、有专业知识但没时间更新的顾问、有产品有客户但内容跟不上的小企业主。选完，产出就是一句话：我服务谁、解决他什么问题。

第 2 步：给他建一份品牌档案
AI 最大的问题不是不会写，是不知道在替谁写。把他过去表现最好的 10–20 条内容找出来，喂给 AI 拆四样东西：反复出现的核心观点、独特的表达习惯、最能打的故事和经历、他反对什么相信什么。没有历史内容的，就访谈式提问：你最想帮谁？你踩过什么坑？你有哪些同行都在说、但你认为不对的观点？产出：一份能反复调用的档案，后面所有内容都从这里长出来，不再每天让 AI 凭空想选题。

第 3 步：搭一张内容地图
定 3–4 个内容支柱（比如：行业洞察 / 个人经历 / 产品与案例 / 反常识观点），每个支柱拆 5–7 个主题，每个主题拆 5–10 个创意，每个创意再拆 4–5 个角度。AI 参考过去 6–12 个月跑得好的内容结构来生成，你负责砍掉不符合定位的。原文说，这样能攒出约半年的内容储备。产出：一张长效选题地图，告别每天问「今天发什么」。

第 4 步：替他写好 10 条推文（这一步是敲门砖）
这 10 条就是你的作品集，配比照着来：2 条个人故事、2 条行业洞察、2 条有逻辑的争议观点、2 条实操经验、2 条立定位的长推。每条认真改到对方几乎能直接发的程度——随便让 AI 吐 10 条垃圾，等于把门敲坏了。

第 5 步：私信直接送成果，不投简历
这招有真实先例。JK Molina 被 Dan Koe 称为开创了推特代笔这行，他当年就是这么私信 Dan Koe 本人的：「我知道你不想让任何人替你写，但我还是替你写了 10 条，你先发发看，喜欢就让我继续写，不喜欢给我个评价就行。」Dan Koe 说，这是他能想到拿下第一个内容客户最好的策略；还有人干脆公开拍视频申请当 Grant Cardone 的创意总监（点名说他的内容还停在 2016 年），Cardone 本人回复说自己正卡着、来联系。
话术照这个结构写：
「嗨，关注你有一段时间了，你的很多观点值得被更多人看到。我猜你最近忙，没太多时间持续更新，所以我研究了你的过往内容和表达风格，直接替你写好了 10 条推文，放在这里了。觉得不错就直接拿去发，不用付费；如果确实有帮助，我们再聊后续合作。」
先给价值，再谈合作。这就是他说的免许可学徒模式：不用等任何人给你许可，先做出成果，拿着成果去敲门。他原话的意思是：真照这么干，他很难相信你一个月拿不下一个客户。

第 6 步：从 10 条推文滚成月费
对方用了、有效果，接下来谈的就不是代写，是整套运营：每周内容计划、发布节奏、数据复盘、把高表现内容改编成长文和 Newsletter。定价阶梯也是他给的参考：首个客户 2500–5000 美元/月先跑通，跑顺了再往 1–3 个客户、月费 1 万到 7.5 万美元的模型走。

为什么说这个窗口是现在？高盛 2023 年 4 月的研究预测，创作者经济会从 2500 亿美元涨到 2027 年的 4800 亿美元。产品做得好不等于市场知道你，创始人懂业务不等于懂表达，这个缺口就是一人创意团队的饭碗。

最后拆穿三条，免得你看完上头：

第一，1 万到 7.5 万美元这个收费，没有市场数据支撑。他没给出成交率、留存率、从业者收入分布，这是他的定价设想，不是普通人照抄就能拿到的收入。能收高价靠的是客户的商业规模和你交付的结果，不是会用几个提示词。

第二，他本人是文中工具 Eden 的创始人，整篇工作流都在给自家工具抬轿子。学他的工作流就行，品牌档案、内容分析、选题规划，用 ChatGPT、Claude 加文档表格一样能搭，不必迷信某个工具。

第三，也是最关键的：AI 写手一点都不稀缺，谁都能让 AI 写 10 条推文。真正稀缺的是判断——这个人该立什么人设、什么热点不能追、一段普通经历怎么讲成有传播力的故事。这才是客户愿意按月付钱的原因。

所以真机会和话术要分开：真机会是帮有真实业务的人，把专业和经历变成持续的内容资产；话术是「学会几个提示词就月入几万」。

今晚就能做的第一步：挑一个你喜欢、但内容明显没做好的创始人，花 30 分钟扒他的账号，找出他最值得传播的 3 个观点，替他写出第 1 条推文。写顺了，补到 10 条，私信发过去。

在 AI 时代，提示词谁都能复制，先做出一份对别人真的有价值的成果，才是你敲得开的那扇门。

[DrFresh @Freshbetterever]

官方稿 https://t.co/c8Ovx2zQwz

[DrFresh @Freshbetterever]

Agent 跑偏了，你连关都关不掉——这件尴尬事，昨天被做成了一个卖钱的产品。
11 月 3 日，CrowdStrike 在 Fal.Con Europe 上，给它今年 9 月就已上线的 Agent 安全平台 Falcon Guardian 加了个新功能：AI Kill Switch。说人话就是给企业里的 AI Agent 装了一个急停按钮：管理员一键下去，跑偏的 Agent 会被掐住、禁止重启、断掉内网访问。今天就能用——但它到底停到哪一步，第二条说清。

[DrFresh @Freshbetterever]

需求有多真？按 CrowdStrike 首席产品官对 Forbes 的说法，急停是 Guardian 每场客户对话都会被问到的功能。官方稿里一位客户安全负责人说得更直白：有了最终决定权，反而敢更快地放行更多 Agent。
同场它还宣布收购做 Agent 可观测的 Fiddler 和做 AI 红队的 Dreadnode（两笔都要到 2027 财年第四季度才完成交割）。一家公司把「看得见、测得出、停得下」一次买齐——这个信号比产品本身更值得看。
你的 Agent，有一个你按得下去的急停键吗？