WEBVTT
Kind: captions
Language: zh

00:00:00.233 --> 00:00:01.400
读一份文档

00:00:01.433 --> 00:00:03.833
ClaudeCode烧掉十几万token

00:00:04.033 --> 00:00:05.633
看个两小时视频

00:00:05.733 --> 00:00:07.600
光字幕又是几万

00:00:07.833 --> 00:00:09.800
今天这个开源工具

00:00:09.800 --> 00:00:12.866
让Google把这些活全部免费干完

00:00:13.100 --> 00:00:14.633
Claude只拿结论

00:00:14.633 --> 00:00:17.066
片尾还有一招免费深度研究

00:00:17.066 --> 00:00:17.866
特别猛

00:00:18.500 --> 00:00:19.500
先看效果

00:00:19.500 --> 00:00:21.733
安装的部分我放最后讲

00:00:21.733 --> 00:00:23.000
就两行命令

00:00:23.000 --> 00:00:26.666
不急假设我现在想吃透MCP

00:00:26.833 --> 00:00:28.300
手上有官方文档

00:00:28.300 --> 00:00:30.533
和一个两小时的课程视频

00:00:30.666 --> 00:00:32.800
我一句话丢给ClaudeCode

00:00:32.900 --> 00:00:35.066
注意看我就说了一句话

00:00:35.133 --> 00:00:37.266
剩下全是它自己在跑

00:00:37.400 --> 00:00:38.866
它建了个笔记本

00:00:38.866 --> 00:00:40.333
把文档丢进去

00:00:40.433 --> 00:00:43.266
现在在等Google把播客生出来

00:00:44.000 --> 00:00:46.133
OK那现在生成好了

00:00:46.200 --> 00:00:48.533
这边会告诉你文件名叫什么

00:00:48.533 --> 00:00:49.666
保存在哪里

00:00:49.700 --> 00:00:52.233
音频的时长文件大小等等

00:00:52.233 --> 00:00:53.233
我来试听一下

00:00:53.233 --> 00:00:54.200
看看效果

00:01:10.666 --> 00:01:13.133
AI主播用中文给我讲MCP

00:01:13.133 --> 00:01:15.933
内容全部来自我喂的官方文档

00:01:16.033 --> 00:01:17.066
不是瞎编的

00:01:17.333 --> 00:01:18.400
重点来了

00:01:18.433 --> 00:01:19.666
这整个过程

00:01:19.666 --> 00:01:22.200
ClaudeCode只发了几行命令

00:01:22.233 --> 00:01:23.533
读文档的累活

00:01:23.533 --> 00:01:25.100
它一个字都没沾

00:01:25.233 --> 00:01:27.000
这笔账等下细算

00:01:27.300 --> 00:01:30.066
你要是把这堆文档直接塞给Claude

00:01:30.200 --> 00:01:32.466
十几万token当场就没了

00:01:32.466 --> 00:01:34.333
而且之后每一轮对话

00:01:34.333 --> 00:01:36.333
都得拖着这坨东西跑

00:01:36.433 --> 00:01:37.866
读文档做播客

00:01:37.866 --> 00:01:38.800
这些重活

00:01:38.800 --> 00:01:41.133
全是NotebookLM免费干的

00:01:41.133 --> 00:01:42.500
它是怎么做到的

00:01:42.500 --> 00:01:45.900
我们从最基本的开始快速交代背景

00:01:46.066 --> 00:01:48.700
NotebookLM是Google的免费AI

00:01:48.700 --> 00:01:50.733
笔记本资料丢进去

00:01:50.733 --> 00:01:53.066
它只根据你给的资料回答

00:01:53.100 --> 00:01:54.433
每句话带引用

00:01:54.433 --> 00:01:55.633
基本不瞎编

00:01:55.866 --> 00:01:57.266
但它只有网页版

00:01:57.266 --> 00:01:58.633
AI碰不了它

00:01:59.000 --> 00:01:59.900
今天的主角

00:01:59.900 --> 00:02:03.200
notebooklm-py社区做的开源项目

00:02:03.233 --> 00:02:05.800
GitHub上17,000多个star

00:02:06.000 --> 00:02:07.933
把它整个搬进了命令行

00:02:08.233 --> 00:02:10.000
环境我先装好了

00:02:10.000 --> 00:02:10.933
直接开始

00:02:10.933 --> 00:02:12.633
那输入这个命令

00:02:12.700 --> 00:02:15.266
它会回给你一个笔记本的id

00:02:15.500 --> 00:02:18.200
注意我结尾加了个--use

00:02:18.333 --> 00:02:19.000
意思是

00:02:19.000 --> 00:02:22.000
把这个新笔记本设成当前笔记本

00:02:22.233 --> 00:02:25.433
后面的命令就都不用再指定id了

00:02:25.633 --> 00:02:27.100
然后开始喂资料

00:02:27.100 --> 00:02:28.800
网页直接给网址

00:02:29.066 --> 00:02:30.933
本地的PDF也行

00:02:31.066 --> 00:02:34.800
我提前从官网存了一份MCP的规格书

00:02:34.800 --> 00:02:36.933
放在当前目录路径

00:02:36.933 --> 00:02:38.133
丢进去就好

00:02:38.300 --> 00:02:41.666
最狠的是YouTube视频直接贴链接

00:02:41.866 --> 00:02:43.333
你想想这个效率

00:02:43.433 --> 00:02:45.466
我以前研究一个新东西

00:02:45.666 --> 00:02:48.200
光把资料看一遍就是一下午

00:02:48.266 --> 00:02:50.466
现在喂完资料直接开问

00:02:50.733 --> 00:02:51.933
看这个回答

00:02:51.933 --> 00:02:54.300
每一段后面都有引用编号

00:02:54.300 --> 00:02:56.500
告诉你这句话是从哪份资料

00:02:56.500 --> 00:02:57.800
哪个位置来的

00:02:58.200 --> 00:03:01.133
到这里你已经会最基本的用法了

00:03:01.300 --> 00:03:03.433
但光拿它问答太浪费了

00:03:03.533 --> 00:03:06.100
它能生出来的东西多到离谱

00:03:06.200 --> 00:03:08.266
下一段全部演给你看

00:03:08.466 --> 00:03:11.300
那这段才是NotebookLM的招牌

00:03:11.633 --> 00:03:12.800
刚才开场那集

00:03:12.800 --> 00:03:14.833
播客命令是这样的

00:03:15.100 --> 00:03:16.500
拆开讲一下

00:03:16.700 --> 00:03:19.700
引号里是你对这集播客的要求

00:03:19.800 --> 00:03:22.433
想让它侧重哪里就写哪里

00:03:22.533 --> 00:03:24.500
--format是对话风格

00:03:24.500 --> 00:03:27.033
默认是deep-dive深度对谈

00:03:27.133 --> 00:03:28.733
我这里选debate

00:03:28.833 --> 00:03:31.066
让两个主播互相抬杠

00:03:31.133 --> 00:03:33.066
听着没那么容易犯困

00:03:33.300 --> 00:03:35.400
--language是简体中文

00:03:35.466 --> 00:03:36.733
支持哪些语言

00:03:36.733 --> 00:03:39.633
你可以用notebooklm language list查

00:03:39.700 --> 00:03:41.733
--wait是让命令停在这里

00:03:41.733 --> 00:03:43.000
等它生成完

00:03:43.466 --> 00:03:44.700
不想傻傻等

00:03:44.700 --> 00:03:47.733
也可以用notebooklm artifact list命令

00:03:47.733 --> 00:03:49.266
查看任务状态

00:03:49.466 --> 00:03:50.833
显示complete

00:03:50.833 --> 00:03:52.300
就可以去下载了

00:03:52.533 --> 00:03:54.700
一份资料喂进去播客

00:03:54.700 --> 00:03:57.000
随时能再生成新的版本

00:03:57.066 --> 00:03:59.700
通勤路上听自己定制的节目

00:03:59.800 --> 00:04:02.066
这体验说实话有点上头

00:04:02.333 --> 00:04:03.400
提醒一句

00:04:03.400 --> 00:04:04.900
免费账号每天的生

00:04:04.900 --> 00:04:06.466
成次数有额度

00:04:06.500 --> 00:04:08.000
学习用足够了

00:04:08.100 --> 00:04:09.866
如果你觉得这个有用

00:04:09.866 --> 00:04:11.066
帮我点个赞

00:04:11.066 --> 00:04:13.266
让更多写代码的人看到

00:04:13.466 --> 00:04:15.800
接着是我最喜欢的功能

00:04:15.900 --> 00:04:16.466
出题目

00:04:16.466 --> 00:04:19.833
来测试学完一个东西到底有没有懂

00:04:19.866 --> 00:04:21.333
做题最诚实

00:04:21.333 --> 00:04:23.600
直接导出成markdown题目

00:04:23.600 --> 00:04:24.866
加标准答案

00:04:24.900 --> 00:04:27.866
丢给团队新人当上手测验都行

00:04:28.100 --> 00:04:31.066
网页版你只能在浏览器里点着玩

00:04:31.066 --> 00:04:32.033
导不出来

00:04:32.200 --> 00:04:34.133
这就是命令行版的好处

00:04:34.266 --> 00:04:36.333
所有东西都变成文件

00:04:36.333 --> 00:04:38.100
能进你自己的流程

00:04:38.266 --> 00:04:39.933
思维导图也一样

00:04:40.133 --> 00:04:41.833
思维导图是同步的

00:04:41.933 --> 00:04:42.833
命令一敲

00:04:42.833 --> 00:04:44.200
结果就回来了

00:04:44.200 --> 00:04:45.466
不用加--wait

00:04:45.833 --> 00:04:48.000
拿到的是结构化的JSON

00:04:48.133 --> 00:04:49.533
你想渲染成图

00:04:49.533 --> 00:04:51.300
想导进笔记软件

00:04:51.300 --> 00:04:52.533
随便你处理

00:04:52.533 --> 00:04:54.433
看不懂JSON也没关系

00:04:54.500 --> 00:04:55.400
丢给Claude

00:04:55.533 --> 00:04:57.466
让它帮你画成图就行

00:04:58.000 --> 00:04:59.400
OK到这里

00:04:59.400 --> 00:05:00.866
你已经完成一半了

00:05:01.033 --> 00:05:03.866
前面这些都还是你手动打命令

00:05:04.100 --> 00:05:06.866
接下来才是今天真正的重头戏

00:05:06.900 --> 00:05:09.066
让ClaudeCode接管这一切

00:05:09.266 --> 00:05:11.133
先做个对照实验

00:05:11.500 --> 00:05:12.333
我现在用的

00:05:12.333 --> 00:05:15.233
是一个还没装外挂的ClaudeCode

00:05:15.466 --> 00:05:17.800
让它总结一个YouTube视频

00:05:18.000 --> 00:05:19.433
它还真干成了

00:05:19.433 --> 00:05:23.066
挺聪明翻出我机器上的yt-dlp

00:05:23.200 --> 00:05:25.533
把字幕扒下来读完了

00:05:25.600 --> 00:05:27.133
但你看这个代价

00:05:27.233 --> 00:05:30.000
看到Messages后面的token数量了吗

00:05:30.000 --> 00:05:31.833
34,000多个token

00:05:31.833 --> 00:05:33.100
就干了一件事

00:05:33.233 --> 00:05:34.500
总结一个视频

00:05:34.666 --> 00:05:35.733
这条路能走

00:05:35.733 --> 00:05:37.000
但又贵又脆

00:05:37.300 --> 00:05:38.833
现在换个姿势

00:05:38.866 --> 00:05:40.133
给它装上外挂

00:05:40.333 --> 00:05:42.400
同样的问题再来一次

00:05:42.466 --> 00:05:44.800
这次我在开头点它的名

00:05:44.933 --> 00:05:46.433
看到区别了吗

00:05:46.500 --> 00:05:48.200
我就多点了它一个名

00:05:48.300 --> 00:05:50.666
它就自己跑完了整套命令

00:05:51.000 --> 00:05:52.933
注意看它调用的每一步

00:05:52.933 --> 00:05:55.000
全是notebooklm字幕

00:05:55.000 --> 00:05:58.066
全文从头到尾没进过它的context

00:05:58.200 --> 00:06:00.333
它只收了NotebookLM递回

00:06:00.333 --> 00:06:01.233
来的答案

00:06:01.800 --> 00:06:05.000
NotebookLM里面其实藏了一个研究代理

00:06:05.200 --> 00:06:06.500
你给它一个题目

00:06:06.500 --> 00:06:08.000
它自己上网搜

00:06:08.000 --> 00:06:09.233
自己筛来源

00:06:09.400 --> 00:06:10.233
网页版里

00:06:10.233 --> 00:06:11.900
这个功能藏得很深

00:06:12.000 --> 00:06:13.400
但是在终端里

00:06:13.466 --> 00:06:15.200
一个命令就可以搞定

00:06:15.300 --> 00:06:16.933
它去网上找了一圈

00:06:17.000 --> 00:06:19.700
把相关的文章文档都翻出来

00:06:19.733 --> 00:06:20.466
跑完之后

00:06:20.466 --> 00:06:22.866
你的笔记本里多了一堆新资料

00:06:22.900 --> 00:06:25.466
这时候再让它生成一份报告

00:06:25.600 --> 00:06:27.066
注意要加--language

00:06:27.200 --> 00:06:29.333
不然报告默认给你英文的

00:06:29.466 --> 00:06:30.900
从我想研究个题目

00:06:30.900 --> 00:06:33.833
到一份带引用的报告躺在我硬盘里

00:06:33.933 --> 00:06:35.433
全程三行命令

00:06:35.433 --> 00:06:36.666
一分钱没花

00:06:36.833 --> 00:06:38.633
这功能单独拿出来

00:06:38.666 --> 00:06:39.833
跟那些一个月

00:06:39.833 --> 00:06:42.266
几十美金的深度研究产品

00:06:42.266 --> 00:06:43.800
干的是同一件事

00:06:44.266 --> 00:06:45.800
再配上ClaudeCode

00:06:45.933 --> 00:06:49.466
你在对话里说一句帮我研究一下XX

00:06:49.466 --> 00:06:51.066
顺便出份报告

00:06:51.066 --> 00:06:52.733
它整套自己跑完

00:06:52.866 --> 00:06:53.633
看到这里

00:06:53.633 --> 00:06:55.000
你应该手痒了

00:06:55.000 --> 00:06:56.666
现在把环境装起来

00:06:56.666 --> 00:06:58.633
很快用uv的话

00:06:58.633 --> 00:06:59.433
一行

00:06:59.600 --> 00:07:03.266
uv是Python圈现在最流行的包管理器

00:07:03.400 --> 00:07:04.266
没装的话

00:07:04.266 --> 00:07:07.400
把开头的uv tool换成pipx install

00:07:07.600 --> 00:07:08.433
包名不变

00:07:08.433 --> 00:07:09.466
效果一样

00:07:09.533 --> 00:07:10.600
装完登录

00:07:10.900 --> 00:07:12.933
诶第一步就翻车

00:07:13.100 --> 00:07:15.633
它说自带的浏览器不存在

00:07:15.700 --> 00:07:17.100
让我再去下载一个

00:07:17.100 --> 00:07:19.066
几百兆的专用浏览器

00:07:19.233 --> 00:07:20.300
别急着下

00:07:20.300 --> 00:07:21.233
加个参数

00:07:21.233 --> 00:07:23.533
直接用你系统里的Chrome就行

00:07:23.866 --> 00:07:27.100
官方文档也推荐Mac用户这么处理

00:07:27.466 --> 00:07:30.333
弹出来的窗口里登你的Google账号

00:07:30.333 --> 00:07:32.900
登进去它自动把凭证存好

00:07:32.900 --> 00:07:33.933
这就完事了

00:07:34.100 --> 00:07:35.533
装好登录

00:07:35.533 --> 00:07:38.300
再加上前面讲的skill install

00:07:38.300 --> 00:07:39.433
一共三步

00:07:39.466 --> 00:07:42.300
你就拥有了开场那套完整工作流

00:07:42.600 --> 00:07:44.400
现在给你一个挑战

00:07:44.433 --> 00:07:45.533
关掉视频

00:07:45.533 --> 00:07:47.200
花5分钟装起来

00:07:47.266 --> 00:07:48.000
把你收藏

00:07:48.000 --> 00:07:49.833
夹里那个一直想看

00:07:49.833 --> 00:07:51.233
但永远没时间看

00:07:51.233 --> 00:07:52.500
的两小时视频

00:07:52.500 --> 00:07:54.733
丢进去生成一集播客

00:07:54.733 --> 00:07:56.100
明天通勤听

00:07:56.200 --> 00:07:57.600
感觉一定很棒

00:07:57.800 --> 00:07:59.600
第一次看我频道的话

00:07:59.600 --> 00:08:01.266
订阅一下不会错过

00:08:01.266 --> 00:08:02.400
我们下期见