Practice实战
AI InAI 实战Practice应用篇
← → 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
01The tools工具05
02What each is for各自用来做什么16
03Real workflows真实工作流21
04Your turn轮到你了35
05Close结尾40
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 / 43
Frame框架
0.1

Part one was the theory. This is the receipts.第一部分是理论。这是实证。

the same four layers, but now with names and price tags还是那四层，但现在有名字和价格
the tools people actually run, and what each one is good at人们真正在用的工具，以及各自擅长什么
real workflows I use, taken apart so you can copy them我实际在用的工作流 workflow，拆开来让你照着抄
nothing here is hypothetical — all of it runs这里没有假设 — 全部都在跑
03 / 43
Frame框架
0.2

Where the leverage actually is.杠杆真正在哪里。

Your procedures你的流程written down, reused, improved — nobody else has these写下来、重复用、持续改 — 别人没有这些
Your material你的材料files, standards, history文件、标准、历史记录
The harness外壳tools, memory, schedule, reach工具、记忆、定时、触达
The model模型same one everybody else has和其他所有人用同一个
the bottom layer is bought. The top three are built — and that is the whole difference最底层是买来的。上面三层是搭起来的 — 差别就在这里
the model is a commodity. What you wrap around it is not模型是商品。套在外面的那层不是
04 / 43
Part One第一部分
01
The tools工具
05 / 43
The tools工具
1.0

Three tiers, and most people only know the first.三个层级，大多数人只知道第一层。

01The chat app聊天应用a window you type into一个让你打字的窗口paste in · copy out粘进去 · 复制出来
02The workspace工作区projects, files, instructions, connectors项目、文件、指令、连接器it reads your material它读你的材料
03The command line命令行agents that run on your machine跑在你机器上的智能体it does the work它干活
most people never leave tier one  ·  the gap between one and three is where everything is大多数人从不离开一层  ·  一层和三层之间的差距才是全部
same models underneath all three. Wildly different reach三层底下用的是同一批模型。能力差距天壤之别
06 / 43
The tools工具
1.1

Tier one: the chat apps.第一层：聊天应用。

ChatGPT, Claude, Gemini — a box you type intoChatGPT、Claude、Gemini — 一个让你打字的框
safe, easy, and it cannot touch anything you own安全、简单，但碰不了你自己的任何东西
you paste things in and copy things out, by hand, every time每次都要手动粘进去、复制出来
this is where 95% of people stop. It is also the weakest tier95% 的人止步于此。这也是最弱的一层
07 / 43
The tools工具
1.2

Tier two: the workspace tools.第二层：工作区工具。

Projects, custom instructions, uploaded files, connectors项目 Projects、自定义指令 custom instructions、上传文件、连接器 connectors
still a window, but it now remembers you and reads your material还是个窗口，但现在它记得你，也会读你的材料
NotebookLM sits here too — answers only from what you uploadNotebookLM 也在这层 — 只从你上传的内容里给答案
same app, ten times the value. Most of it is free and switched off by default同一个应用，价值翻十倍。大部分功能免费，默认关闭
08 / 43
The tools工具
1.3

Tier three: the command line.第三层：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
it stops being a chat and starts being a worker它不再是对话，变成了一个干活的
09 / 43
The tools工具
1.3

What changes at tier three.第三层有什么不同。

it reads and writes your actual files, not copies you pasted它读写你的真实文件，不是你粘贴进去的副本
it runs for many minutes on one instruction一条指令可以跑很多分钟
it can call other software, and check its own work它能调用其他软件，也能检查自己的工作
it can be scheduled, so it runs when you are asleep可以定时，让它在你睡觉时运行
the jump from tier two to tier three is bigger than the jump from one to two从二层跳到三层，比从一层跳到二层跨度更大
10 / 43
The tools工具
1.4

Claude Code.Claude Code。

Anthropic’s coding agent, lives in your terminalAnthropic 的编程智能体 AI agent，住在你的终端里
works on a folder — plans, edits, runs, checks, repeats在一个文件夹里工作 — 规划、编辑、运行、检查、循环
reads a house-rules file so it knows your standards读取规则文件，了解你的标准
strongest of the four at long, multi-step work四个里面最擅长长流程、多步骤的工作
a folder of files describes a case bundle or a client folder as well as it describes software一个文件夹可以描述一批案卷或客户档案，和描述软件一样顺手
11 / 43
The tools工具
1.5

Codex.Codex。

OpenAI’s equivalent, also terminal-nativeOpenAI 的对应产品，同样原生在终端
sandboxed by default — it asks before it touches anything real默认沙箱隔离 — 碰任何真实内容前会先问你
reads the same kind of house-rules file, and can call outside tools读同样类型的规则文件，也能调用外部工具
included with a ChatGPT subscription rather than metered separately包含在 ChatGPT 订阅里，不单独计费
if you already pay OpenAI, this is the one you already own and are not using如果你已经付费用 OpenAI，这个你已经有了，只是没用
12 / 43
The tools工具
1.6

OpenClaw.OpenClaw。

not a coding tool — a personal assistant that lives on your own machine不是编程工具 — 是住在你自己机器上的个人助理
reaches your messaging apps, calendar, browser, files and email能访问你的消息应用、日历、浏览器、文件和邮件
runs on a schedule, so work happens without you asking定时运行，工作在你不开口的情况下自动发生
keeps memory in plain files you can read and correct记忆 memory存在普通文件里，你可以读取和修改
the difference: you do not open it. It comes to you区别在于：你不用打开它，它主动来找你
13 / 43
The tools工具
1.7

Hermes.Hermes。

open-source, from Nous Research开源，来自 Nous Research
the harness is not fixed — it writes its own procedures from experience外壳 harness不是固定的 — 它从经验中写出自己的流程
improves them mid-task and keeps them across sessions在任务中途改进流程，并在会话间保留
searches its own past work before starting new work开始新工作前，先搜索自己过去的工作
the direction of travel: harnesses that accumulate instead of resetting趋势方向：外壳不再重置，而是不断积累
14 / 43
The tools工具
1.8

The honest summary.实话实说。

Claude Code
long multi-step work长流程多步骤工作
a folder of files一个文件夹
strongest at planning规划能力最强
metered or subscription按量计费或订阅制
Codex
same shape of job同类型的工作
sandboxed by default默认沙箱隔离
asks before it acts行动前先询问
comes with ChatGPT包含在 ChatGPT 里
OpenClaw
personal assistant个人助理
messages, calendar, browser消息、日历、浏览器
runs on a schedule定时运行
self-hosted自托管
Hermes
learns as it goes边干边学
writes its own procedures自己写流程
open source开源
earliest of the four四个里最早出现的
you do not need all four. Pick the one whose shape matches your week你不需要全部四个。选那个形态最符合你这周工作的
15 / 43
Part Two第二部分
02
What each is for各自用来做什么
16 / 43
What each is for各自用来做什么
2.1

Match the tool to the shape of the job.根据工作的形态选工具。

one question, one answer一个问题，一个答案the chat app聊天应用
must know my files and my style必须了解我的文件和我的风格a workspace with your material in it一个装了你材料的工作区
two hundred files, same operation两百个文件，同一个操作a coding agent编程智能体
must happen every Monday at 9必须每周一九点执行a scheduled assistant定时助理
nobody sells the tool I need没人卖我需要的工具a coding agent, building it for you编程智能体，帮你搭
17 / 43
What each is for各自用来做什么
2.2

The test that tells you which tier you need.判断你需要哪一层的测试。

one answer, and you act on it — tier one is fine一个答案，直接用 — 一层够了
needs your documents or your standards — tier two需要你的文件或你的标准 — 二层
touches many files, or must happen on a schedule — tier three涉及大量文件，或必须定时执行 — 三层
most people reach for tier one for a tier-three job, then conclude AI is overrated大多数人拿三层的活去用一层工具，然后得出 AI 被高估的结论
18 / 43
What each is for各自用来做什么
2.3

What a coding agent is actually for, if you do not code.如果你不写代码，编程智能体 AI agent到底能用来干什么。

reading a folder of documents and pulling out the facts读一批文件，把事实提取出来
converting things — format to format, language to language, mess to table做转换 — 格式转格式、语言转语言、乱数据转表格
building you a small private tool nobody sells给你搭一个市面上没人卖的小私有工具
doing the same operation to two hundred files对两百个文件做同一个操作
hand it a problem you have no method for — it works one out, within its limits丢一个你没有方法的问题给它 — 它会在能力范围内想出一个办法
you are not learning to code. You are learning to describe the job你不是在学写代码。你是在学如何描述这个活
19 / 43
What each is for各自用来做什么
2.4

What a scheduled assistant is for.定时助理用来干什么。

work that should happen whether or not you remember it不管你记不记得都应该发生的工作
reports that assemble themselves before you wake up在你醒来之前自动拼好的报告
watching something and only speaking when it matters持续监视某件事，只在有意义时才开口
the value is not speed. It is that it never forgets and never gets bored价值不在于速度，在于它永远不会忘事，也永远不会厌倦
20 / 43
Part Three第三部分
03
Real workflows真实工作流
21 / 43
Real workflows真实工作流
3.0

Everything after this runs in my own setup.接下来的一切都在我自己的系统里跑。

names, brands and clients are stripped out名字、品牌和客户都已去除
the mechanism is exactly as it is机制完全照原样
none of it was built in one sitting — all of it started small没有一个是一口气搭完的 — 全都从小处开始
copy the shape, not the specifics抄结构，不要抄细节
22 / 43
Real workflows真实工作流
3.1

The morning brief.早间简报。

SCHEDULED定时任务
06:00a schedule fires, nobody is awake定时任务触发，没人醒着
→
gather收集prices, headlines, calendar价格、头条、日历
→
write生成a model turns it into a short brief模型把它变成一份简短汇报
→
send发送it arrives on my phone发到我手机上
runs every day whether I remember it or not  ·  this is the whole shape of a scheduled assistant不管我记不记得都每天跑  ·  这就是定时助理的全部形态
23 / 43
Real workflows真实工作流
3.1

What that one actually replaced.这个实际上替代了什么。

twenty to thirty minutes of hunting news and prices, every morning每天早上二三十分钟找新闻和价格
the days I forgot, which were most of them我忘记的那些天，也就是大多数时候
built in an afternoon, running for six months. When one data source died it was swapped in under an hour, and the habit never broke一个下午搭起来，跑了六个月。某个数据源挂掉时，不到一小时换好，习惯从没断过
24 / 43
Real workflows真实工作流
3.2

Long video into a written summary.长视频转成文字摘要。

PIPELINE流水线
a 2-hour video一段两小时的视频no transcript没有文字稿
→
pull the audio提取音频one command line tool一个命令行工具
→
transcribe转录a speech model, pennies语音模型，几分钱
→
summarise摘要into the format I always want转成我一直想要的格式
→
publish发布posted where I will read it发到我会看到的地方
a recorded meeting, a lecture or a client call works exactly the same way录制的会议、讲座或客户电话，用同样的方式处理
25 / 43
Real workflows真实工作流
3.2

Why this one is worth copying first.为什么这个值得第一个抄。

it is the easiest big win on this list这是这份清单里最容易拿到的大收益
every step is a free or near-free tool每一步用的都是免费或近似免费的工具
a two-hour recording becomes a three-minute read两小时录音变成三分钟阅读
same pipeline works on a recorded meeting, a lecture, or a client call同一条流水线可以处理录制的会议、讲座或客户电话
26 / 43
Real workflows真实工作流
3.3

A weekly report that builds itself.自动生成的周报。

BROWSER + SCRIPT浏览器 + 脚本
log in登录a browser agent does the clicking浏览器智能体负责点击
→
pull the numbers拉取数据the same report, every week每周同一份报告
→
do the maths做计算a script, not a model — maths must be exact用脚本，不用模型 — 数学必须精确
→
render渲染one locked design file一个锁定的设计文件
→
deliver交付posted, with a record kept发出去，保留记录
note step three: the model never touches the arithmetic. That is deliberate注意第三步：模型从不碰算术，这是故意的
27 / 43
Real workflows真实工作流
3.3

The part people miss.大家通常忽略的部分。

the hard bit was never the numbers — it was logging in and clicking难的从来不是数字 — 而是登录和点击
a browser agent does the clicking; a script does the maths浏览器智能体 AI agent负责点击；脚本负责计算
the design was locked once in a style file, so every week matches设计在样式文件里锁定一次，之后每周都一致
roughly five manual steps a week, gone. The output looks better than the version I made by hand每周大约五个手动步骤消失了。输出比我手做的版本更好看
28 / 43
Real workflows真实工作流
3.4

Reading a pile of documents nobody wants to read.读一堆没人想读的文件。

FAN-OUT并行分工
81 emails81 封邮件pulled straight from the mailbox直接从邮箱拉取
→
4 chat exports4 份聊天记录导出plus two long contracts加上两份长合同
→
read in parallel并行阅读several agents, one per source多个智能体，每个来源一个
→
cross-reference交叉比对dates, claims, contradictions日期、论断、矛盾
→
one cited brief一份带引用的简报every claim points at its source每个论断都指向来源
the citation rule is what makes it usable  ·  without it you have a confident guess引用规则是让它可用的关键  ·  没有它你只有一个自信的猜测
29 / 43
Real workflows真实工作流
3.4

What that looked like in practice.实际上是什么样的。

eighty-one emails, four chat exports and two long contracts八十一封邮件、四份聊天记录导出和两份长合同
pulled, read, cross-referenced and turned into one cited brief拉取、阅读、交叉比对，整合成一份带引用的简报
roughly three hours, against a professional quoting twenty大约三小时，而专业人士报价二十小时
the rule that made it safe: every claim points back to the document it came from让它可靠的规则：每个论断都指向来源文件
30 / 43
Real workflows真实工作流
3.5

Two models arguing, so I do not have to guess.两个模型互相争，我就不用猜了。

a decision worth being right about一个值得做对的决策
Model A模型 Aargues one way持一种立场
Model B模型 Bargues the other持另一种立场
What I actually read我实际读的where they disagree它们分歧在哪
one model agrees with you too easily  ·  two of them argue, and the argument is the value一个模型太容易顺着你  ·  两个互相争，争论本身就是价值
31 / 43
Real workflows真实工作流
3.5

Why bother with two.为什么要用两个。

one model agrees with you too easily一个模型太容易跟你同意
a second, different model disagrees in useful places第二个不同的模型在有价值的地方提出异议
I read the disagreement, not the summary我读分歧，不读摘要
used weekly on decisions where being wrong is expensive每周用于代价高昂的决策
32 / 43
Real workflows真实工作流
3.6

A small tool nobody sells.一个没人卖的小工具。

I needed honest pricing for a collectables market with filthy data我需要一个数据质量极差的收藏品市场的真实定价
wrote the rules once — sold listings only, never mix grades, match the variant exactly规则只写一次 — 只看成交记录、不混级别、精确匹配版本
now it is a saved procedure I call by name, and it answers the same way every time现在它是一个按名字调用的存储流程，每次给出相同的答案
this is the difference between asking a question and owning a tool这就是问一个问题和拥有一个工具的区别
33 / 43
Real workflows真实工作流
3.7

The pattern under all of them.所有这些背后的规律。

Every one of them每一个
started as one boring task都从一个无聊的任务开始
ran manually first先手动跑
got written down, then automated写下来，再自动化
None of them没有一个
was planned as a system是被规划成系统的
took more than a day to start花了超过一天才开始
needed me to write code by hand需要我手写代码
The maths rule数学规则
models draft, describe and sort模型负责起草、描述和整理
scripts do arithmetic脚本做算术
never let a model add up your money永远不要让模型帮你算钱
34 / 43
Part Four第四部分
04
Your turn轮到你了
35 / 43
Your turn轮到你了
4.1

Pick the task you already resent.选那个你已经在抵触的任务。

it repeats, it is boring, and it is not judgement它是重复的、无聊的，不需要判断力
you already know the steps — you just do them by hand你已经知道步骤 — 只是在手动做
you would not be sad if you never did it again这件事你永远不做也不会可惜
if you cannot describe the steps to a person, do not hand it to a machine如果你没办法把步骤讲清楚给一个人听，就别交给机器
36 / 43
Your turn轮到你了
4.2

The order that actually works.真正有效的顺序。

01Do it by hand, with the chat app open手动做，把聊天应用开着learn where it actually helps搞清楚它真正能帮上忙的地方this week这周
02Write the steps down把步骤写下来the moment you repeat yourself在你重复做同一件事的那一刻week two第二周
03Hand the written steps to a tool把写好的步骤交给工具now it is repeatable现在它可以重复执行了week three第三周
04Put it on a schedule把它设成定时only once it is boring and stable只在它已经无聊且稳定之后later than you want比你想要的时间晚
skipping straight to four is the single most common way this fails直接跳到第四步是最常见的失败方式
37 / 43
Your turn轮到你了
4.3

The rule I break most and regret every time.我最常违反、每次都后悔的规则。

run it by hand for two weeks before you automate it先手动跑两周，再自动化
if the process does not survive being manual, automating it just breaks it faster and more quietly如果流程手动跑都撑不住，自动化只会让它崩得更快、更悄无声息
every one of my workflows that died, died from skipping this我所有死掉的工作流 workflow，都死于跳过这一步
38 / 43
Your turn轮到你了
4.4

Start smaller than feels worth it.从小到感觉不值得的程度开始。

one task, one tool, this week这周，一个任务，一个工具
not a system. Not a plan. One task不是一套系统，不是一个计划，就一个任务
the second one is ten times easier, and you will not believe how much第二个会容易十倍，你现在不会信，但以后会
everything in Part Three started as something that felt too small to bother with第三部分的所有东西，最初都小到感觉不值得动手
39 / 43
Part Five第五部分
05
Close结尾
40 / 43
Close结尾
5.1

What to take away.带走什么。

the model is the commodity — the wrapping is the leverage模型是商品 — 外壳是杠杆
tier three is where the real work happens, and it is not hard to reach三层才是真正干活的地方，而且并不难到达
your own material and your own procedures are the moat你自己的材料和你自己的流程才是护城河
start with the task you resent从你最抵触的任务开始
41 / 43
Close结尾
5.2

The default assumption.默认假设。

if you find yourself wondering whether AI could do something, it probably can如果你开始想 AI 能不能做某件事，它大概能
ask it the moment you think of it, not later想到的时候马上问，别拖
it will tell you what it needs from you它会告诉你它需要你提供什么
the cost of asking is nothing. The cost of assuming it cannot is another year of doing it by hand问一下的成本是零。假设它不行的代价是又一年手动去做
42 / 43
Questions.提问。
Part three — Make It Yoursconnectors, MCP and skills →第三部分 — 把它变成你的连接器、MCP 和技能 →
jrforeal.com
01 / 43
ask ai

which ai do you use?

asked once. you can change it any time.

cancel
ask ai
~/practice $
change ai esc
中文
Contents目录click a slide to jump  ·  Esc to close点击任意页跳转  ·  按 Esc 关闭