Concepts概念
How AIAI 是如何Actually Works运作的
← → to move  ·  C for contents  ·  L for 中文  ·  F for fullscreen← → 翻页  ·  C 打开目录  ·  L 切换 English  ·  F 全屏Tap the right to go forward, the left edge to go back
Swipe works too  ·  the grid icon opens the contents点击右侧前进，点击左边缘后退
也可以滑动  ·  点网格图标打开目录
jrforeal.com
Contents目录

Contents目录

00Frame框架01
01What it is它是什么06
02The harness外壳23
03Your own knowledge base你自己的知识库54
04Agentic AI智能体化 AI63
05In practice实际操作72
06Close结尾80
press C at any point to jump between slides任何时候按 C 都可以跳转到其他页
02 / 85
Frame框架
0.1

What this session covers.本次课程涵盖的内容。

how the pieces fit together各个部分如何组合在一起
what each piece can and cannot do每个部分能做什么、不能做什么
enough structure to judge what is worth your time足够的框架，让你判断哪些值得花时间
03 / 85
Frame框架
0.2

Why we start with concepts.为什么从概念开始讲。

the tools change every few months工具每隔几个月就会更新
what sits underneath them does not底层原理不会变
once you know the structure, a new tool is just a new interface一旦掌握结构，新工具只是新界面
04 / 85
Frame框架
0.3

The four layers, bottom to top.四个层级，从底层到顶层。

Agent智能体 AI agenta harness given a goal, allowed to loop被赋予目标、可以循环执行的外壳
Harness外壳 harnessdocuments, memory, tools, instructions文件、记忆、工具、指令
Model模型produces text生成文字
Compute算力data centre, chips, tokens数据中心、芯片、token
every AI product you have heard of is one of these four things你听说过的每一款 AI 产品，都属于这四类之一
05 / 85
Part One第一部分
01
What it is它是什么
06 / 85
What it is它是什么
1.1

It runs in a data centre, not on your laptop.它运行在数据中心，而不是你的电脑上。

Your laptop你的电脑sends the text发送文字
→←
DATA CENTRE数据中心
thousands of chips, working at once数千块芯片同时运作
nothing is computed on your machine  ·  the round trip is why it takes a few seconds你的机器上不进行任何运算 · 来回传输就是为什么需要几秒钟
07 / 85
What it is它是什么
1.1

Tokens are the unit of measurement.Token 是计量单位。

Thanksforyourinvoicelastweek
7 tokens, not 6 words  ·  what you send and what it returns are both counted
谢谢你上周寄来的发票
10 个汉字，7 个 token  ·  你发出去的和它返回的都要算
a token is roughly three quarters of a word一个 token 大约相当于四分之三个英文单词
08 / 85
What it is它是什么
1.1

What that means for your bill.这对你账单意味着什么。

a one-line question一行问题~20
a page of notes一页笔记~700
a 20-page report一份 20 页报告~14,000
a 300-page bundle一份 300 页文件包~200,000
roughly how many tokens each one costs to read  ·  your monthly plan is an allowance of these大概读取各项内容所需的 token 数量  ·  你的月度计划就是这些的额度
nothing is broken when you hit a limit — you have spent the allowance触及限额并非出错 — 只是额度用完了
09 / 85
What it is它是什么
1.2

It has a knowledge cut-off — and a patch for it.它有知识截止日期—也有解决办法。

it learned from text up to a fixed date, and nothing after它从固定日期之前的文本中学习，此后的内容一概不知
web search fills that gap — most tools now look things up when they need to网络搜索填补了这个空缺—大多数工具现在会在需要时自动查询
but it only searches when it decides to, and you cannot see when it did not但它只在自己判断需要时才搜索，你看不到它什么时候没有搜索
and it has never known anything about your business, searchable or not而且它对你的业务一无所知，不管是否可以搜索到
the cut-off matters less than it used to. What it never had is your material截止日期的影响比以前小了。它从来不知道的是你自己的资料
10 / 85
What it is它是什么
1.3

It predicts the next piece of text.它预测下一段文本。

Thanks for your email. I will▮
get71%
check17%
reply8%
look4%
pick one  ·  append  ·  repeat
谢谢你的邮件，我会▮
尽快71%
查17%
回复8%
看4%
选一个  ·  接上去  ·  重复
there is no database being consulted没有数据库被查询
11 / 85
What it is它是什么
1.3

It produces what is likely, not what is true.它生成的是最可能的答案，而不是真实的答案。

likely and true overlap most of the time大多数时候，可能的和真实的是一致的
they separate on specifics: names, numbers, citations, dates在具体细节上会出现偏差：姓名、数字、引用、日期
12 / 85
What it is它是什么
1.4

It can be confidently wrong.它可能一本正经地讲错。

Reliable可靠
structure结构
tone and register语气与文体
summarising what you gave it总结你提供的内容
rewriting and shortening改写与精简
Unreliable不可靠 verify需核实
quotes and sources引用与来源
figures, dates, reference numbers数字、日期、参考编号
anything recent任何近期信息
anything specific to you任何涉及你具体情况的内容
a wrong answer looks exactly as convincing as a right one. This cannot be fixed, only managed — and the fix is you checking错误的答案看起来和正确的一样令人信服。这无法修复，只能管控—解决办法是你自己核查
13 / 85
What it is它是什么
1.5

The context window is its working memory.上下文窗口 context window是它的工作记忆。

instructions指令
your documents你的文件
the conversation so far目前为止的对话
free空闲
everything it can see at once  ·  measured in tokens它一次能看到的所有内容 · 以token计量
when it fills, the earliest material has to go somewhere当它填满时，最早的内容必须被处理掉
14 / 85
What it is它是什么
1.5

Compaction: older material is summarised, not deleted.压缩 compaction：旧内容被摘要处理，而非删除。

Before压缩前
full detail — window nearly full完整细节—窗口接近满载
→
After compaction压缩后
summary摘要
gist kept — exact wording lost要点保留—原文措辞丢失
this is why it remembers the topic but forgets a figure you gave an hour ago这就是为什么它记得话题，却忘了你一小时前提供的数字
15 / 85
What it is它是什么
1.5

Why long chats still get worse — and what to do.为什么长对话仍然会越来越差—以及应对方法。

each round of compaction loses a little more precision, so it drifts每次压缩 compaction都会损失一些精度，因此它会逐渐偏离
Start fresh — a new chat with the key material pasted back in重新开始—开一个新对话，把关键材料重新粘贴进去
Keep facts in files, not in the chat — the chat is working memory, the file is the record把事实存在文件里，不要存在对话里—对话是工作记忆 memory，文件才是记录
Use memory and Projects — so the standing facts reload every time instead of being retold使用记忆和 Projects 功能—让固定事实每次自动加载，而不是反复重述
a firmer instruction does not fix drift. Re-supplying the material does更强硬的指令无法修正偏差，重新提供材料才行
16 / 85
What it is它是什么
1.6

Two speeds: answers now, or thinks first.两种速度：立即回答，或先思考再回答。

Instant — replies straight away. ChatGPT Instant, Gemini Flash即时型—立即回复。ChatGPT Instant、Gemini Flash
Thinking — works through steps first. ChatGPT Thinking, Claude extended thinking, Gemini Pro思考型—先逐步推理再回答。ChatGPT Thinking、Claude extended thinking、Gemini Pro
Pro — the longest and most expensive tier, for research-grade work专业型—最慢也最贵的层级，适合研究级工作
thinking is slower and costs more, and is better on anything multi-step思考型更慢、更贵，但在多步骤任务上表现更好
the version numbers change every few months. These three tiers do not版本号每隔几个月就会变，但这三个层级不会
17 / 85
What it is它是什么
1.7

A handful of labs make the frontier models.少数几个实验室开发前沿模型。

OpenAI (ChatGPT), Anthropic (Claude), Google (Gemini), and several othersOpenAI（ChatGPT）、Anthropic（Claude）、Google（Gemini）以及其他几家
the ranking changes every few months排名每隔几个月就会变动
pick one and learn it properly — they behave the same way选一个认真学透—它们的运作方式是一样的
18 / 85
What it is它是什么
1.7

Closed models and open models.闭源模型与开源模型。

Closed闭源
runs on the lab's servers运行在研究机构的服务器上
strongest available当前最强的模型
your text leaves the building你的文本会传出公司
ChatGPT OpenAI  ·  Claude Anthropic  ·  Gemini Google
Open开源
the model file is published模型文件已公开发布
weaker, and you host it能力较弱，需自行部署
runs on your machine — nothing leaves运行在你的机器上—数据不外流
Llama Meta  ·  Qwen Alibaba  ·  DeepSeek  ·  Mistral  ·  Gemma Google
open models are weaker, but the data never leaves the building开源模型能力较弱，但数据不会离开你的内部环境
19 / 85
What it is它是什么
1.8

Text is one modality among several.文本只是多种模态 modality之一。

Text文本
today's subject今天的主题
Image图像
generate and read生成与识别
Audio音频
speech both ways双向语音
Video视频
generate生成
same underlying method, different output底层方法相同，输出形式不同
20 / 85
What it is它是什么
1.8

Vision: it can read what it is shown.视觉 vision：它能读懂所看到的内容。

photo · scan · screenshot照片 · 扫描件 · 截图
→
supplier供应商Tan Brothers Sdn Bhd
invoice no发票编号INV-20418
date日期14 July7月14日
amount金额RM 8,420.00
it reads the picture and hands back the fields  ·  handwriting included它读取图片并返回各字段  ·  包括手写内容
this is the bridge between your paper files and everything else here这是你纸质文件与本课所有内容之间的桥梁
21 / 85
What it is它是什么
1.9

Subscription versus API.订阅方案与 API。

Subscription订阅
fixed monthly fee固定月费
use the app使用应用程序
capped usage有使用上限
where you are now你目前所在的阶段
API
pay per token按 token 计费
no app — software only没有应用程序—仅限软件接入
uncapped, metered无上限，按量计费
matters once you automate一旦自动化，这点就很重要
most people never touch the API. Even once you automate things, the tools usually sign in with your account (OAuth) and run on your subscription大多数人从不接触 API。即使你自动化了流程，工具通常也会用你的账户（OAuth）登录并在你的订阅下运行
22 / 85
Part Two第二部分
02
The harness外壳 Harness
23 / 85
The harness外壳
2.0

Think of it as a car.把它想象成一辆汽车。

ENGINE引擎
Engine引擎the model — enormous power, but on its own it only spins模型—能力强大，但单独运行只会空转
The car around it围绕它的车身the harness — chassis, fuel line, wheels, dashboard外壳—车架、油管、车轮、仪表盘
Fuel燃料your documents and instructions你的文件和指令
Self-driving mode自动驾驶模式an agent — you name a destination instead of steering一个智能体（AI agent）—你指定目的地，而不是亲自驾驶
tyres and steering are still car parts, not agents. What makes it an agent is who is driving, not which part you added轮胎和方向盘仍是汽车零件，不是智能体 AI agent。让它成为智能体的，是谁在驾驶，而不是你加了哪个零件
24 / 85
The harness外壳
2.1

The model alone can do nothing useful.模型本身什么有用的事都做不了。

stripped back, it is a pile of maths — numbers and rules, nothing more剥去外壳，它只是一堆数学 — 数字和规则，仅此而已
it cannot open a file, search, or remember you它无法打开文件、搜索，也无法记住你
it cannot send anything anywhere它无法发送任何内容到任何地方
text in, text out, nothing else. Everything else you have seen it do was something around it文字输入，文字输出，仅此而已。你见过它做的其他一切，都是外围系统在发挥作用
25 / 85
The harness外壳
2.1

The harness is everything wrapped around the model.外壳 harness（harness）是包裹在模型外面的一切。

HARNESS外壳 HARNESS
your documents你的文件
MODEL模型text in · text out文字输入·文字输出
conversation history对话历史
tools & connectors工具&连接器
standing instructions常驻指令
ChatGPT the app is a harness — the model sits inside itChatGPT 这款应用就是一个外壳（harness）—模型在它里面
26 / 85
The harness外壳
2.1

Same model, different harness, very different results.相同的模型，不同的外壳 harness，结果大相径庭。

put a Ferrari engine in a Myvi frame and it is still a Ferrari engine把 Ferrari 引擎装进 Myvi 车壳，它仍然是 Ferrari 引擎
it will not drive like a Ferrari, because the car around it is not one但它开起来不像 Ferrari，因为外壳不是
same the other way round: an ordinary model in a good harness beats a great model in a bad one反过来也一样：普通模型配好外壳，胜过强大模型配差外壳
most of what people call “AI got better this year” is the harness getting better人们所说的“AI 今年进步了”，大多数是外壳在进步
27 / 85
The harness外壳
2.2

System prompt: instructions it gets before you speak.系统提示词 system prompt：它在你开口之前收到的指令。

written and hardcoded by the company that makes the product由产品公司编写并硬编码进去
you cannot see it, edit it or remove it你看不到它，也无法编辑或删除它
it outranks anything you type — where you conflict with it, it wins它的优先级高于你输入的任何内容—当你与它冲突时，它胜出
your custom instructions sit underneath it, not above it你的自定义指令 custom instructions在它下面，而不是上面
this is why some requests are refused however you word them这就是为什么有些请求无论你怎么措辞都会被拒绝
28 / 85
The harness外壳
2.2

This is one layer of the guardrails.这是护栏 guardrails的其中一层。

Safety training — taught during training to refuse certain things安全训练—在训练阶段教会模型拒绝某些内容
The system prompt — standing rules written by the vendor系统提示词 system prompt—由供应商编写的固定规则
Output filters — a separate check on the answer before you see it输出过滤器—在你看到回答之前单独进行的一道审查
Account policy — usage terms, rate limits, and bans账号政策—使用条款、频率限制和封禁规定
four layers, not one. It is why a refusal does not move however you rephrase it四层，不是一层。这就是为什么无论你怎么换说法，拒绝都不会改变
29 / 85
The harness外壳
2.2

Custom instructions: your standing brief.自定义指令 custom instructions：你的固定简报。

who you are, what you do, how you want output returned你是谁、你做什么、你希望输出以什么形式返回
applied automatically to every chat自动应用于每次对话
written once只需写一次
removes the need to re-explain yourself every time省去每次重新解释自己情况的麻烦
30 / 85
The harness外壳
2.3

Prompting is briefing.写提示词 prompting就是做简报。

Role — who it should act as角色—它应该扮演什么角色
Task — what to produce任务—要输出什么
Context — the material to work from背景—工作所依据的材料
Format — how the answer should be structured格式—回答应如何组织结构
a vague brief produces vague work, exactly as with a pupil模糊的简报产生模糊的结果，和对待学生一样
31 / 85
The harness外壳
2.3

Examples work better than instructions.举例比给指令更有效。

describing the standard you want is slow and imprecise用语言描述你想要的标准既慢又不精确
showing one good example of the output is fast and exact展示一个好的输出样例，既快又准确
32 / 85
The harness外壳
2.3

The single most useful instruction.最有用的一条指令。

“Use only the documents I have given you. If it is not there, say so.”“只使用我提供的文件。如果信息不在其中，请说明。”
this converts a guessing machine into a reading machine这把一台猜测机变成了一台阅读机
33 / 85
The harness外壳
2.4

Giving it your documents: retrieval.提供你的文件：检索 retrieval。

your files你的文件
indexed已建立索引
relevant pages pulled提取相关页面
model answers from those pages模型根据这些页面作答
the common name for this is RAG — retrieval augmented generation这通常被称为 RAG—检索增强生成
34 / 85
The harness外壳
2.4

Retrieval matters more than model choice.检索 retrieval比选择模型更重要。

the model does not know your files, your customers or your standards模型不了解你的文件、你的客户或你的标准
retrieval is the only way it ever will检索是让它了解这些的唯一途径
your own material is the advantage; the model is a commodity你自己的资料才是优势，模型只是通用工具
35 / 85
The harness外壳
2.4

Retrieval is not the same as training.检索 retrieval与训练不是一回事。

Training / fine-tuning — changes the model itself: slow, costly, rarely needed训练 / 微调 fine-tuning—改变模型本身：速度慢、成本高、很少需要
Retrieval — leaves the model alone and hands it the right pages at the right moment检索—不改动模型，在恰当时机将正确的内容交给它
almost every real use case is retrieval, not training几乎所有实际应用场景都是检索，而非训练
36 / 85
The harness外壳
2.4

Grounded tools restrict it to your documents only.有依据的工具将它限制在你的文件范围内。

NotebookLM is the clearest exampleNotebookLM 是最典型的例子
it answers from what you upload and from nothing else它只从你上传的内容中作答，别无其他
each answer points back to the page it came from每个答案都指向其来源页面
the closest available thing to a system that cannot invent这是目前最接近无法凭空捏造内容的系统
37 / 85
The harness外壳
2.5

Memory: what it keeps between conversations.记忆 memory：对话之间保留的内容。

facts it stores about you across chats跨对话保存的关于你的信息
you can view, edit and delete it你可以查看、编辑和删除这些信息
different from the context window, which covers one conversation only与上下文窗口 context window不同，上下文窗口只覆盖单次对话
38 / 85
The harness外壳
2.6

Projects: a workspace with permanent files.项目 Projects：带有持久文件的工作空间。

upload your templates and standards once一次性上传你的模板和标准
set the instructions for that project once为该项目设置一次指令
every chat started inside it begins with all of that loaded在其中开始的每次对话都会加载所有这些内容
this is the main structure you will end up using这是你最终会主要使用的结构
39 / 85
The harness外壳
2.7

Tools: how the model acts on the world.工具：模型如何对外部世界采取行动。

search the web搜索网络
run a calculation执行计算
read or write a file读取或写入文件
call another piece of software调用其他软件
the model asks for a tool; the harness actually runs it模型请求调用工具；外壳 harness负责实际执行
40 / 85
The harness外壳
2.7

How a tool call works.工具调用的工作原理。

you ask你来问
model requests a tool模型请求调用工具
harness runs it外壳执行工具
result returned结果返回
model continues模型继续运行
this loop is the basis of everything agentic这个循环是所有智能体化 agentic功能的基础
41 / 85
The harness外壳
2.8

Connectors: prebuilt links to services you use.连接器 connectors：与你所用服务的预置链接。

Google Drive, Gmail, Calendar, Outlook, SharePoint and similarGoogle Drive、Gmail、Calendar、Outlook、SharePoint 及类似服务
you authorise access once你只需授权一次
the AI can then read from them directlyAI 就可以直接从这些来源读取内容
this is how it reaches your real work instead of text you paste in这样它就能直接访问你的实际工作内容，而不是你粘贴进来的文字
42 / 85
The harness外壳
2.9

MCP: a common standard for connecting tools.MCP：连接工具的通用标准。

ChatGPT
Claude
Codex
MCPone standard plug一个标准插口
Drive
Calendar
your database你的数据库
a standard socket, rather than a different adapter for every appliance统一的标准接口，而非每种设备各用一种不同的转接头
43 / 85
The harness外壳
2.9

What MCP changes.MCP 带来的变化。

a tool built once works across ChatGPT, Claude and others一次构建的工具可跨 ChatGPT、Claude 及其他平台使用
the range of available tools grows without each vendor's permission可用工具的范围在无需各厂商逐一授权的情况下持续扩展
you do not need to build these — you need to know they exist你不需要自己搭建这些 — 你只需知道它们的存在
44 / 85
The harness外壳
2.10

Two ways to reach the model.访问模型的两种方式。

App应用程序
a window and buttons有界面和按钮
safe defaults安全的默认设置
limited reach功能受限
most people stop here大多数人止步于此
Command line命令行
reads and writes your files读写你的文件
runs long tasks执行长时间任务
chains steps unattended无人值守地串联执行步骤
where agents actually live智能体真正发挥作用的地方
the same model underneath both两者底层使用的是同一个模型
45 / 85
The harness外壳
2.10

Why anyone uses the command line.为什么有人会使用命令行 command line。

bash
~/invoices $ read every PDF in this folder and list who owes me what读取此文件夹中的所有 PDF，列出谁欠我多少▮
reading 38 files...正在读取 38 个文件……
Tan Brothers — RM 8,420 — 21 days overdueTan Brothers — RM 8,420 — 逾期 21 天
Lim Hardware  — RM 2,180 — 6 days overdueLim Hardware  — RM 2,180 — 逾期 6 天
written to owed.csv已写入 owed.csv
plain English in, real work out  ·  it touches your actual files输入普通文字，输出实际成果  ·  它直接操作你的真实文件
no window, no buttons — you type what you want and it goes and does it没有窗口，没有按钮 — 你输入想要的，它就去执行
46 / 85
The harness外壳
2.11

Coding agents: Codex, Claude Code and similar.编程智能体 AI agent：Codex、Claude Code 及类似工具。

built for software, but not only useful for software为软件开发而生，但不仅限于软件用途
they operate on a folder of files它们在一个文件夹内运行
they plan, act, check their own work, and repeat它们会规划、执行、自我检查，然后循环往复
the most capable form of harness generally available today目前普遍可用的功能最强的外壳 harness形式
47 / 85
The harness外壳
2.11

Why anyone should care about a coding tool.为什么普通人也应该关注编程工具。

“a folder of files” describes a project folder as well as software“一个文件夹”既可以描述项目文件夹，也可以描述软件
it can process hundreds of documents from one instruction它可以根据一条指令处理数百份文件
it can build small tools on request — a tracker, a converter, a report它可以按需构建小工具 — 追踪器、转换器、报告
you are not learning to code; you are describing what you want你不是在学写代码，你是在描述你想要什么
48 / 85
The harness外壳
2.12

Skills: saved procedures it follows on command.技能 skill：按指令执行的已保存流程。

a written set of steps for a task you repeat针对重复性任务编写的一套步骤
stored once, called by name存储一次，按名称调用
the same output standard every time每次输出标准一致
the difference between explaining a job each time and handing over an SOP每次重新解释工作与直接交给它一份标准操作流程之间的区别
49 / 85
The harness外壳
2.12

What makes a good skill.什么构成一个好的技能 skill。

a task done the same way repeatedly以相同方式反复执行的任务
with a defined output有明确的输出结果
where consistency matters more than originality一致性比独创性更重要的场景
your standard replies and recurring reports are exactly this shape你的标准回复和定期报告正好符合这种形式
50 / 85
The harness外壳
2.13

Stop briefing it. Let it interview you.别再给它做简报了。让它来采访你。

Reverse prompting反向提问
you ask it what it needs to know你问它需要了解什么
it interviews you about your work它来向你询问你的工作
Alex Finn — broad discoveryAlex Finn—广泛探索
Grill me质问我
a saved skill that attacks a plan一个用来审视计划的已保存技能
one question at a time, until the holes show一次一个问题，直到漏洞显现
adversarial depth对抗性深度
same move both ways round: the AI pulls the brief out of you, instead of waiting for one两种方向的相同操作：AI从你身上提取简报，而不是等你提供
it feels back to front. It works because the AI knows what it is missing better than you know what to tell it这感觉有点反常。它之所以有效，是因为 AI 比你更清楚自己缺少什么信息
51 / 85
The harness外壳
2.13

What that looks like.这看起来是什么样的。

“Based on what you know about me and my work, what else should I tell you so you can help me faster?”“根据你对我和我工作的了解，我还应该告诉你什么，才能让你更快地帮到我？”
“Grill me on this plan. One question at a time. Find the holes before I commit.”“就这个计划质问我。一次一个问题。在我决定之前找出漏洞。”
most people find the questions expose gaps they did not know were there大多数人发现这些问题暴露出他们之前没有意识到的信息空缺
52 / 85
The harness外壳
2.14

Agents that learn from use.从使用中学习的智能体 AI agent。

most tools start every session from zero and forget what worked大多数工具每次会话都从零开始，不记得什么方法有效
Hermes Agent — Nous Research, open-source, released February 2026Hermes Agent — Nous Research，开源，2026 年 2 月发布
it runs a learning loop: writes its own skills from experience, improves them while working, and keeps them across sessions它运行一个学习循环：从经验中写出自己的技能 skill，在工作中持续改进，并跨会话保留这些技能
it can also search back through its own past conversations它还可以回溯检索自己过去的对话记录
the direction of travel: the harness stops being fixed and starts accumulating发展方向：外壳 harness不再是固定的，而是开始积累经验
53 / 85
Part Three第三部分
03
Your own knowledge base你自己的知识库
54 / 85
Your knowledge base你的知识库
3.1

The model is generic. Your material is not.模型是通用的。你的素材不是。

everyone has access to the same models每个人都能使用同样的模型
nobody else has your files, your history, your standards没有人拥有你的文件、你的历史记录、你的标准
the value sits in what you give it价值在于你提供给它的内容
55 / 85
Your knowledge base你的知识库
3.2

A second brain: your knowledge in one place.第二大脑 second brain：将你的知识集中在一处。

notes, templates, procedures, decisions笔记、模板、流程、决策
plain files, plain text普通文件，纯文本
readable by you and by any AI tool你和任何 AI 工具都可以读取
not an app — a folder and a habit不是一个应用 — 而是一个文件夹和一种习惯
56 / 85
Your knowledge base你的知识库
3.2

Why plain text files.为什么选择纯文本文件。

no lock-in to one vendor不被任何单一厂商锁定
readable by every tool, current and future当前及未来的所有工具均可读取
searchable, versioned, permanent可搜索、可版本管理、永久保存
Markdown is the usual format — a text file with light formattingMarkdown 是常用格式 — 带有简单排版的文本文件
57 / 85
Your knowledge base你的知识库
3.2

The version worth copying: Karpathy’s LLM wiki.值得效仿的版本：Karpathy’s LLM wiki。

you drop in the raw material — notes, articles, meeting scraps你放入原始素材 — 笔记、文章、会议碎片
an AI agent turns it into tidy linked pages: one per client, topic or decision智能体 AI agent将其整理成有序的关联页面：每个客户、主题或决策各占一页
it writes and maintains the wiki. You read it and correct it它编写并维护知识库。你阅读并纠正
plain markdown files in a folder, so you can always see what it wrote普通 Markdown 文件存在文件夹里，你随时可以查看它写了什么
named after Andrej Karpathy, who published the pattern. The point is the filing gets done by the machine — that is the part everyone gives up on以 Andrej Karpathy 命名，他公开了这个方法。关键在于归档由机器完成 — 这正是所有人都放弃的环节
58 / 85
Your knowledge base你的知识库
3.2

Why that version works.这个版本为何有效。

the effort that kills every note system is the filing, and it is gone毁掉所有笔记系统的是归档工作，而它现在消失了
you can read and audit every page, unlike a black-box search index你可以阅读和审查每一页，不像黑盒搜索索引
it improves as you feed it, without you tidying anything随着你的输入它会不断改善，无需你整理任何东西
start with one folder and one topic. It is a habit, not a project从一个文件夹、一个主题开始。这是习惯，不是项目
59 / 85
Your knowledge base你的知识库
3.3

What goes in it, and what does not.哪些内容放进去，哪些不放。

team-knowledge/
templates/ proposals · contracts · replies提案·合同·回复
house-style.md how we word things我们的措辞方式
procedures/ enquiry to delivery从询盘到交付
decisions/ what we chose, and why我们的选择及理由
reference/ facts looked up repeatedly反复查阅的参考事实
no live customer data不含实时客户数据
keep the structure separate from the confidential content将结构与保密内容分开存放
60 / 85
Your knowledge base你的知识库
3.4

Structure beats volume.结构胜过数量。

one file, one topic一个文件，一个主题
clear titles清晰的标题
files linked to related files文件与相关文件相互链接
a small tidy library beats a large messy one, for you and for the machine整洁的小型资料库胜过杂乱的大型资料库，对你和对机器都是如此
61 / 85
Your knowledge base你的知识库
3.5

Why this compounds.为什么这会产生复利效应。

every answer improves because the source improved每个答案都因为来源的改善而变得更好
new staff read the same files新员工阅读同样的文件
a new AI tool is useful on day one because the material already exists新的 AI 工具从第一天起就能发挥作用，因为素材已经存在
the files outlast whichever tool happens to be best this year这些文件的寿命超过任何一款今年最好用的工具
62 / 85
Part Four第四部分
04
Agentic AI智能体化 Agentic AI
63 / 85
Agentic AI智能体化 AI
4.1

Three levels of use.三个使用层级。

Chat对话
you ask, it answers你提问，它作答
you drive every step每一步都由你主导
Workflow工作流
a fixed sequence固定流程
same path every time每次走同一条路径
Agent智能体 AI agent
you set a goal你设定目标
it decides the steps它决定步骤
64 / 85
Agentic AI智能体化 AI
4.2

What makes something an agent.什么使一个工具成为智能体 AI agent（AI agent）。

a goal rather than a question一个目标，而非一个问题
tools it is permitted to use被允许使用的工具
permission to loop — act, observe, decide, act again循环执行的权限 — 行动、观察、决策、再行动
a condition for stopping停止的条件
65 / 85
Agentic AI智能体化 AI
4.2

The loop, plainly.简单说，就是循环。

plan规划
act行动
check核查
adjust调整
until直到
done完成
the same thing a junior does with a file — that is the entire idea就像初级员工处理一份文件时做的事 — 这就是全部的核心思路
66 / 85
Agentic AI智能体化 AI
4.3

Autonomy is a dial, not a switch.自主程度 autonomy是一个旋钮，不是一个开关。

1Suggests, you execute提出建议，由你执行drafts the reply — you read it and press send起草回复—你阅读后点击发送
2Acts, then asks先行动，再请示prepares the document and waits for your sign-off准备好文件，等待你审批
3Acts, reports after先行动，事后汇报updates the tracker, reports at 6pm what it did更新跟踪表，晚上6点汇报操作内容
4Acts unattended无人监督下自主行动emails the customer directly, with nobody reading it first直接发邮件给客户，无人事先审阅
go only as far down as you would let a first-week hire go unsupervised授权范围以你对刚入职员工的信任程度为限
67 / 85
Agentic AI智能体化 AI
4.4

Several agents can work together.多个智能体 AI agent可以协同工作。

one plans, several work in parallel, one checks一个规划，多个并行执行，一个负责检查
suited to volume — many documents, many checks适合大批量任务 — 大量文件，大量检查
more capacity, and more ways to go wrong更强的处理能力，也有更多出错的可能
68 / 85
Agentic AI智能体化 AI
4.5

Where agents fail.智能体 AI agent在哪里会出错。

one wrong step early, then ten more built confidently on top of it早期一个错误步骤，之后十个步骤都信心满满地建立在它之上
work that looks finished but is not看起来完成了但实际上没有完成的工作
an unclear goal produces unclear work目标不清晰，产出也不清晰
they do not notice when they are lost它们不会察觉自己已经偏离方向
69 / 85
Agentic AI智能体化 AI
4.5

Cost and time behave differently.成本和时间的表现方式不同。

an agent may run for many minutes and use a great deal一个智能体 AI agent可能运行数分钟，消耗大量资源
it is making hundreds of model calls, not one它在进行数百次模型调用，而不是一次
this is why agentic tools are priced differently这就是为什么智能体化 agentic工具的定价方式不同
70 / 85
Agentic AI智能体化 AI
4.6

The supervision rule.监督原则。

match oversight to consequence监督力度与后果影响相匹配
reversible and internal — let it run可逆且内部使用 — 放手让它运行
anything that leaves the office — you read it first任何离开公司的内容—你先过目
never let an agent be the last set of eyes on something a customer receives永远不要让智能体 AI agent成为客户收到的内容的最后一道审核
71 / 85
Part Five第五部分
05
In practice实际操作
72 / 85
In practice实际操作
5.1

Confidentiality: what actually happens.保密性：实际发生了什么。

whatever you type is sent to a company’s servers and processed there你输入的内容会发送到服务商的服务器上处理
on free and personal plans it may be used to train the model — there is a setting to turn that off, and you should在免费和个人计划中，可能会用于训练模型 — 有一个设置可以关闭，你应该关掉
on business and API plans it is not used for training at all在商业和 API 计划中，完全不用于训练
decide what you will and will not paste in before you are in the middle of a job在开始工作之前，先决定哪些内容可以粘贴，哪些不可以
73 / 85
In practice实际操作
5.1

The habit that solves most of it.能解决大部分问题的习惯。

it does not need the real name, ID number or account number to do the work它完成工作并不需要真实姓名、身份证号或账号
swap them for initials or XXX — you get the same draft back用缩写或 XXX 代替 — 你得到的草稿一样好
put the real details in yourself at the end最后自己填入真实信息
if you do only one thing from this section, do this one如果这部分你只做一件事，就做这件
74 / 85
In practice实际操作
5.2

Always check these three things.始终核查这三点。

numbers — figures, dates, totals数字 — 数据、日期、合计
names — people, companies, places名称 — 人名、公司、地点
sources — anything it says it is quoting来源 — 它声称引用的任何内容
writing style and structure you can trust. Facts you cannot写作风格和结构可以信任。事实不能
75 / 85
In practice实际操作
5.2

Never let it answer from memory.永远不要让它凭记忆 memory作答。

asking “what does the law say about X” invites it to make something up问 “法律对 X 有何规定” 会引导它编造内容
giving it the document and asking “what does this say about X” does not给它文件并问 “这里对 X 说了什么” 则不会
same question, completely different reliability同样的问题，可靠性截然不同
no source in front of it means the answer is a guess, however confident it sounds没有原始资料在前，答案就是猜测，无论听起来多么肯定
76 / 85
In practice实际操作
5.3

You still sign it.最终还是由你署名。

whatever goes out has your name on it, not the AI’s发出去的内容署的是你的名字，不是 AI 的
“the AI wrote it” will not help you if it is wrong“AI 写的” 在出错时帮不了你
so read it before it leaves所以在发出前先阅读
treat it as a fast assistant, not as the person who is responsible把它当作快速助手，而不是负责任的人
77 / 85
In practice实际操作
5.4

What to hand over and what to keep.哪些可以交出去，哪些要自己保留。

Automate可自动化
repetitive重复性任务
clearly defined定义清晰的任务
internal内部任务
low consequence影响较小的任务
Keep manual保持人工处理 you由你来做
judgement判断
advice建议
negotiation谈判
anything final任何最终决定
the win is removing repetition, not removing thinking收益在于消除重复，而不是消除思考
78 / 85
In practice实际操作
5.5

Not everything needs AI.不是所有事情都需要AI。

some jobs are just admin, not thinking — those need automation, not AI有些工作只是行政事务，不需要思考 — 那需要自动化，而不是 AI
a rule, a template or a spreadsheet is often the right answer规则、模板或表格往往才是正确答案
AI on top of a broken process just breaks it faster在有缺陷的流程上加 AI，只会让它坏得更快
knowing which is which will save you more time than anything else here分清哪些是哪些，比这里任何其他方法都更省时间
79 / 85
Part Six第六部分
06
Close结尾
80 / 85
Close结尾
6.1

The whole picture.全局来看。

You你supply the judgement提供判断
Agent智能体 AI agentruns the loop运行循环
Harness外壳 harnessgives it context and tools提供上下文和工具
Model on compute运行在算力上的模型produces the text生成文字
81 / 85
Close结尾
6.2

What actually determines the result.真正决定结果的是什么。

the material you give it你给它的素材
the clarity of the instruction指令的清晰程度
your judgement on the way out你在输出环节的判断
none of that depends on which tool you choose这些都与你选择哪个工具无关
82 / 85
Close结尾
6.3

The order to learn it.学习的顺序。

use one tool properly把一个工具用好
put your own material into it把你自己的素材放进去
write down the procedures you repeat把你重复执行的流程写下来
automate only what is already stable只自动化已经稳定的内容
hands-on from here从现在开始动手实践
83 / 85
Close结尾
6.4

The default assumption to leave with.带走的默认假设。

if you wonder whether AI could do something — it probably can如果你在想AI能不能做某件事—它很可能能做到
do not sit on the question. Ask the AI itself, the moment you think of it不要把问题搁置。一有想法，马上去问AI
it will tell you whether it is possible, and what it would need from you它会告诉你是否可行，以及它需要你提供什么
the cost of asking is nothing. The cost of assuming it cannot is another year by hand问一下的代价是零。假设它做不到的代价，是再多一年手工操作
84 / 85
Questions.提问。
jrforeal.com
01 / 85
中文
Contents目录click a slide to jump  ·  Esc to close点击任意页跳转  ·  按 Esc 关闭