20260801-34 | Give Me 10 Mins and I'll Save You Millions of Claude Tokens
來源:Youtube | 建立:2026-08-01T21:44:42 | HTML:2026-08-01T21:47:31
開啟原始影片 note.md transcript.txt transcript.vtt

影片筆記:Give Me 10 Mins and I'll Save You Millions of Claude Tokens

YouTube 影片框會固定在左上方;點擊右側逐字稿時間戳可跳到對應時間。

一句話總結

講者透過解析 Claude 的 Prompt Caching(提示詞快取)機制,說明如何透過理解快取層級、TTL 規則及善用專案管理,在單日節省數千萬 Tokens,並強調根據自身需求篩選資訊的重要性。

核心重點

Prompt Caching 的成本優勢:快取內的 Tokens 僅收取正常輸入成本的 10%,Anthropic 高度重視快取命中率。

TTL(生存時間)差異

快取結構層級

快取創建與讀取

影響快取的因素

實用工具與習慣

個人學習策略

詳細大綱

一、Prompt Caching 效益與成本結構

二、快取機制與 TTL 規則

三、快取結構層級解析

四、快取創建(Create)與讀取(Read)

五、影響快取的因素與誤區

六、實用工具與習慣

不要暫停太久(超過 1 小時請開啟新會話)。

切換任務時重新開始。

使用專案管理大型文件。

七、結尾:資訊篩選與個人需求

工具 / 模型 / 名詞整理

操作流程整理

監控與追蹤

專案管理優化

會話維護與移交

複製當前摘要。

清除會話。

貼上摘要以模擬「會話移交」。

指令操作

模型切換注意事項

資訊篩選

值得注意的限制或風險

TTL 失效風險

模型切換成本

系統提示詞變更

快取創建成本

大型文件處理

Opus Plan 模式限制

逐字稿辨識疑點

可延伸追問

如何具體設定 API 與 Subagents 的 TTL 以平衡成本與便利性?

"SEVs" 在 Anthropic 內部或特定情境下的具體定義為何?

"Session Handoff Skill" 的具體操作步驟與最佳實踐為何?

對於重度使用 API 的用戶,有哪些進階的快取優化策略?

如何準確識別並避免 "Opus Plan" 模式下的模型切換風險?

"Token Dashboard" 的具體功能與使用介面為何?

如何評估將大型文件放入 Projects 的具體效益與成本?

對於非重度 API 用戶,如何有效篩選資訊以避免資訊過載?

逐字稿時間軸

右側可一路往下捲;左側影片框會固定。點擊時間戳會讓左側影片跳到對應秒數。

00:00:00.000 → 00:00:06.720
So look at this, on this day I saved 91 million tokens because of cash read, and in the past week I've saved over 300 million tokens because of it.
00:00:06.840 → 00:00:12.680
Now don't freak out, this isn't anything that you have to go change, this is happening automatically if you are using Claude Code or Claude.
00:00:12.880 → 00:00:17.460
And I know that the concept of prompt caching might seem a little bit overwhelming, but today I'm going to make it as simple as possible,
00:00:17.720 → 00:00:22.420
and only really tell you what you need to know in order to make sure that you are saving your session limits and saving tokens.
00:00:22.780 → 00:00:26.900
I'll also give you guys this entire token dashboard for free so you can actually start tracking your tokens a little bit better.
00:00:26.900 → 00:00:31.180
Anyway, so let's talk about prompt caching, why your sessions burn out, and how to stop it.
00:00:31.300 → 00:00:33.020
So what does caching actually cost you?
00:00:33.360 → 00:00:37.180
Well, cached tokens only cost you 10% of normal input.
00:00:37.360 → 00:00:39.720
So all the tokens that are getting cached are saving you a ton of money.
00:00:40.000 → 00:00:44.560
So if we go back to this example, on this day when I had 91 million tokens cached,
00:00:45.060 → 00:00:48.600
that costed me only as if I was processing about 9 million of those tokens.
00:00:48.820 → 00:00:51.260
The cache window on a Cloud subscription is an hour,
00:00:51.420 → 00:00:54.780
meaning if you're working with Cloud Code and you don't touch it for an hour,
00:00:54.780 → 00:00:56.120
and then you send another message,
00:00:56.420 → 00:00:58.800
everything in that session gets uncashed.
00:00:58.960 → 00:01:01.200
So if you leave a session sitting for an hour or longer,
00:01:01.620 → 00:01:03.180
then you're gonna pay more for it.
00:01:03.300 → 00:01:05.400
And if you're using Claude via API or subagents,
00:01:05.500 → 00:01:08.180
then the TTL or the time to live is only five minutes.
00:01:08.340 → 00:01:09.160
You can change that,
00:01:09.200 → 00:01:10.120
but it's just a little bit more expensive.
00:01:10.260 → 00:01:11.440
You could bump it up to an hour if you want.
00:01:11.440 → 00:01:13.560
But for Claude code inside of your terminal
00:01:13.560 → 00:01:15.780
or your extension, whatever it is, that's an hour.
00:01:16.020 → 00:01:17.700
And now here's a quote from Thorik from Anthropic.
00:01:18.100 → 00:01:19.500
He said that we actually run alerts
00:01:19.500 → 00:01:20.620
on our prompt cache hit rate
00:01:20.620 → 00:01:22.500
and declare SEVs if they're too low.
00:01:22.640 → 00:01:23.560
So basically them saying,
00:01:23.560 → 00:01:25.240
we take this stuff really, really seriously.
00:01:25.780 → 00:01:27.900
And if we see that the hit rate isn't very high
00:01:27.900 → 00:01:30.240
for users' cloud code caching,
00:01:30.520 → 00:01:32.180
then we do something about it immediately.
00:01:32.480 → 00:01:33.320
And that's very nice of them,
00:01:33.380 → 00:01:35.420
but also, of course, it benefits themselves
00:01:35.420 → 00:01:37.440
because with a high cache hit rate,
00:01:37.800 → 00:01:38.720
cloud code feels faster,
00:01:39.180 → 00:01:40.420
their serving cost is lower,
00:01:40.780 → 00:01:42.500
subscription limits feel more generous,
00:01:42.700 → 00:01:44.000
you know, because you're using less,
00:01:44.300 → 00:01:46.160
and long coding sessions stay practical.
00:01:46.560 → 00:01:48.200
And then if you have low cache hit rate,
00:01:48.340 → 00:01:49.140
this is what happens.
00:01:49.180 → 00:01:51.220
And obviously, it's just a lose-lose for everybody.
00:01:51.220 → 00:01:56.900
and that's why I said like prompt caching can get very very complex and if you want to check out
00:01:56.900 → 00:02:01.960
more then I'll link this article right here which Thorik really goes into some depth here but if
00:02:01.960 → 00:02:05.240
you read this at least when I did I was like okay this is a little bit overwhelming I have a feeling
00:02:05.240 → 00:02:09.440
I don't actually need to know all of this but I do need to know at least a little bit at least you
00:02:09.440 → 00:02:14.020
know the 80 20 of prompt caching so that I can get the most out of my session limits and that's
00:02:14.020 → 00:02:18.280
what I'm going to break down today so let's take a look at an example of how this actually grows
00:02:18.280 → 00:02:24.040
so by default when you shoot off a message to claude there's going to be some information that
00:02:24.040 → 00:02:27.240
needs to be cached right away and actually let me just switch back to one of thorax graphics real
00:02:27.240 → 00:02:32.220
quick you can see here that we have the base system instructions get globally cached we have
00:02:32.220 → 00:02:37.460
tools like read write bash grab glob globally cached we have per memory or sorry per project
00:02:37.460 → 00:02:42.280
things like claude.md and memory and that gets cached per project we've got session state and
00:02:42.280 → 00:02:45.620
and then we have user messages which grow each turn.
00:02:45.620 → 00:02:48.480
So now that we take this into context
00:02:48.480 → 00:02:49.660
and we flip back over here,
00:02:49.660 → 00:02:50.660
this is what it looks like.
00:02:50.660 → 00:02:52.920
This is an example where we have four turns.
00:02:52.920 → 00:02:54.140
So on turn one, there's no cache.
00:02:54.140 → 00:02:55.940
Basically we're matching on a prefix.
00:02:55.940 → 00:02:58.840
So don't really have to worry about what that means,
00:02:58.840 → 00:03:00.060
but I might mention that later.
00:03:00.060 → 00:03:02.300
So anyways, on turn one, there's nothing, right?
00:03:02.300 → 00:03:03.840
We're opening up a fresh session.
00:03:03.840 → 00:03:06.360
We load in the system prompt, the project context,
00:03:06.360 → 00:03:08.160
and we shoot off our first message.
00:03:08.160 → 00:03:11.860
And all of this is kind of in this like brown highlight border,
00:03:11.860 → 00:03:13.860
which means that this is new
00:03:13.860 → 00:03:15.360
and it has to be fully processed
00:03:15.360 → 00:03:17.800
and it's being written to the cache here.
00:03:17.800 → 00:03:19.880
So before I continue down this graphic,
00:03:19.880 → 00:03:21.500
in this dashboard you can see
00:03:21.500 → 00:03:23.880
that we have the difference between cache create
00:03:25.100 → 00:03:27.420
So on these days you can see what are my input tokens,
00:03:27.420 → 00:03:29.580
my output tokens and my cache create.
00:03:29.580 → 00:03:31.740
And then over here you can see my daily cache reads.
00:03:31.740 → 00:03:33.220
And just a quick explanation,
00:03:33.220 → 00:03:35.860
a cache create is writing something into cache
00:03:35.860 → 00:03:36.940
for the first time.
00:03:36.940 → 00:03:39.740
It's a one-time cost and it pays off the next turn
00:03:39.740 → 00:03:41.340
unless of course everything gets uncached
00:03:41.340 → 00:03:43.700
and the cache read is tokens that Claude reused
00:03:43.700 → 00:03:45.560
from a cache like your Claude.md
00:03:45.560 → 00:03:46.300
or some of the files
00:03:46.300 → 00:03:48.720
or some of the global system instructions.
00:03:48.720 → 00:03:50.900
And these are the things that are 10 times cheaper
00:03:50.900 → 00:03:52.120
than fresh input.
00:03:52.120 → 00:03:53.960
So anyways, on turn two,
00:03:53.960 → 00:03:57.320
given that we're within that one hour TTL window,
00:03:57.320 → 00:04:00.140
everything here is already in context, so it's cached.
00:04:00.140 → 00:04:02.240
And then all that Claude actually has to process
00:04:02.240 → 00:04:05.360
for the first time is reply one and message two,
00:04:05.360 → 00:04:06.440
and it caches that.
00:04:06.440 → 00:04:08.880
So then down here in turn three, all of that's cached,
00:04:08.880 → 00:04:11.400
and we are bumping up a reply and a message
00:04:11.400 → 00:04:13.680
and those are the things that only get processed each time.
00:04:13.840 → 00:04:15.380
But if we waited an hour
00:04:15.380 → 00:04:16.740
and then we sent another message
00:04:16.740 → 00:04:18.640
or if we changed the system prompt,
00:04:18.960 → 00:04:20.660
then everything from the very beginning
00:04:20.660 → 00:04:22.320
has to get fully recached.
00:04:22.580 → 00:04:25.600
So imagine if you were on message like 16
00:04:25.600 → 00:04:27.340
and you're way, way, way over here on the right
00:04:27.340 → 00:04:29.680
and you change the system prompt or you wait an hour,
00:04:30.340 → 00:04:31.940
then everything getting recached
00:04:31.940 → 00:04:34.720
is going to be a pretty expensive move that you just made.
00:04:34.960 → 00:04:36.920
So anyways, once again, we have the system layer,
00:04:37.020 → 00:04:38.440
the project layer and the conversation layer,
00:04:38.440 → 00:04:43.700
the system layer has instructions, tool definitions, output style, and here's where it might break.
00:04:44.000 → 00:04:49.480
the project level or the project layer has Cloud.md, memory, and rules, and then here's when that might break.
00:04:49.760 → 00:04:52.980
And then we have, of course, the conversation, which is just like the replies and the messages,
00:04:53.320 → 00:04:56.660
which gets recached every time, but that's how it should be.
00:04:56.800 → 00:04:59.700
So here's where there's been some confusion among the community.
00:04:59.840 → 00:05:05.500
So how long does the cache snapshot live, which is kind of called the TTL, the time to live?
00:05:05.500 → 00:05:10.860
so on your cloud subscription you have an hour by default because it uses your subscription
00:05:10.860 → 00:05:15.580
but if you go over that weekly limit and you are now playing in your extra usage territory where
00:05:15.580 → 00:05:21.860
you are paying per token api then by default that will be five minutes which is very dangerous if
00:05:21.860 → 00:05:25.960
you're managing multiple sessions and you're constantly recaching everything because five
00:05:25.960 → 00:05:29.380
minutes is passing you got to be careful about that and people were kind of suspicious i don't
00:05:29.380 → 00:05:32.880
know if you remember like a month or so ago when everyone was complaining about their cloud
00:05:32.880 → 00:05:38.080
subscriptions, how quick they were eating it up. People thought maybe that they switched the cache
00:05:38.080 → 00:05:42.860
TTL from an hour to five minutes without like saying anything to anybody. It turns out they
00:05:42.860 → 00:05:47.340
didn't. So it is an hour, but that's just like, you know, there was a lot of confusion around that.
00:05:47.340 → 00:05:52.360
And I get why, because honestly, it's not super clear. Like if you're on an API, you have five
00:05:52.360 → 00:05:57.040
minutes by default, but you can increase the cost and you can do an hour. And then your sub agents
00:05:57.040 → 00:06:00.740
on any plan are going to be five minutes. And for some reason, all of this is documented about
00:06:00.740 → 00:06:03.080
ClawedCode and the API, which are two very different things.
00:06:03.520 → 00:06:07.860
But the Clawed.ai, like on the web, we don't know exactly how that works.
00:06:08.160 → 00:06:10.460
At least I haven't found documentation on that exact.
00:06:10.740 → 00:06:15.280
I'm assuming it's the same as your subscription, but I don't know 100% for fact.
00:06:15.500 → 00:06:18.220
Anyways, three habits that cover 95% of people.
00:06:18.700 → 00:06:19.540
Don't pause too long.
00:06:19.540 → 00:06:24.900
So if you've gone over an hour on a session, just hand it off to a new session.
00:06:25.120 → 00:06:26.920
Obviously, start fresh when you switch tasks.
00:06:27.280 → 00:06:30.340
So do a slash compact, which will break the cache, or do a slash clear.
00:06:30.340 → 00:06:32.340
或是用我的SessionHandoffSkill
00:06:32.340 → 00:06:34.340
我會加上的作品
00:06:34.340 → 00:06:36.340
所以我會加上的TokenDashboard
00:06:36.340 → 00:06:37.340
GithubRepo和這個Skill
00:06:37.340 → 00:06:39.340
會在我的FreeschoolCommunity
00:06:39.340 → 00:06:42.340
但基本上,我會說
00:06:42.340 → 00:06:43.340
我會有這個Project
00:06:43.340 → 00:06:45.340
幫助我建立這個HTML
00:06:45.340 → 00:06:47.340
它們有205,000TOKENS
00:06:47.340 → 00:06:49.340
如果我來做SessionHandoff
00:06:49.340 → 00:06:51.340
這基本上整理我們做的
00:06:51.340 → 00:06:53.340
所有的重要招待
00:06:53.340 → 00:06:54.340
所有的開招待
00:06:54.340 → 00:06:55.340
所有的開招待
00:06:55.340 → 00:06:58.340
然後我可以直接
00:06:58.340 → 00:07:03.540
that summary do a slash clear and then keep going and it feels like i haven't actually lost anything
00:07:03.540 → 00:07:07.780
so that has been basically my replacement for doing slash compact i've just enjoyed doing this
00:07:07.780 → 00:07:11.540
better and sometimes the compact takes a long time this typically doesn't take anywhere over
00:07:11.540 → 00:07:16.260
a minute there you go so that is my session handoff i do a slash copy and then i just go
00:07:16.260 → 00:07:20.420
ahead and clear that paste it in hit enter and now i'm basically right back where i was and then
00:07:20.420 → 00:07:24.660
this last one is for if you're using claude chat specifically if you're going to be pasting big
00:07:24.660 → 00:07:28.180
documents in there you're probably better off doing a project because like i said i don't
00:07:28.180 → 00:07:31.620
I don't know exactly how the caching works in Cloud Chat,
00:07:31.620 → 00:07:35.180
but we do have some confidence in saying that projects,
00:07:35.180 → 00:07:37.220
those files are cached a little bit differently
00:07:37.220 → 00:07:40.260
and probably more optimized for storing a bunch of documents
00:07:40.260 → 00:07:42.260
compared to just dropping them into your Cloud Chat.
00:07:42.645 → 00:07:43.745
So keep it alive, keep it focused,
00:07:43.745 → 00:07:45.685
and start fresh when you switch.
00:07:45.685 → 00:07:46.805
Now there's a few other things
00:07:46.805 → 00:07:48.505
that were a little bit confusing to me
00:07:48.505 → 00:07:50.885
as far as like what breaks the cache.
00:07:50.885 → 00:07:53.185
So the first one is if you switch the model.
00:07:53.185 → 00:07:55.205
So you know if you're in here
00:07:55.205 → 00:07:56.445
and you're talking to Cloud,
00:07:56.445 → 00:07:57.645
hello,hello,hello
00:07:57.645 → 00:07:59.325
and then you go in here and you do a slash model
00:07:59.325 → 00:08:00.825
and you actually switch the model
00:08:00.825 → 00:08:02.325
that's going to recache everything
00:08:02.325 → 00:08:03.405
because if you remember earlier
00:08:03.405 → 00:08:04.565
I said it's prefix matching
00:08:04.565 → 00:08:06.165
which I'm not going to dive into right now
00:08:06.165 → 00:08:07.665
but if you switch the model
00:08:07.665 → 00:08:09.965
then you are switching essentially the prefix
00:08:09.965 → 00:08:11.825
and it can't match on that same cache
00:08:11.825 → 00:08:13.205
so if you switch the model
00:08:13.205 → 00:08:15.065
you are recaching everything
00:08:15.065 → 00:08:16.585
now I do want to apologize for something here
00:08:16.585 → 00:08:19.725
because if you do model opus plan
00:08:19.725 → 00:08:21.165
which is something that I've shown before
00:08:21.165 → 00:08:23.265
in like token hacks videos
00:08:23.265 → 00:08:25.385
this basically means it uses opus for plan mode
00:08:25.385 → 00:08:28.525
and then it switches to Sonnet for the execution.
00:08:28.525 → 00:08:30.365
But if you do that, just keep in mind,
00:08:30.365 → 00:08:31.545
that's actually gonna break the cache
00:08:31.545 → 00:08:33.505
because you're switching model halfway through.
00:08:33.505 → 00:08:35.925
So right here, you can see each model has its own cache.
00:08:35.925 → 00:08:38.065
Switching with model means the next request
00:08:38.065 → 00:08:40.705
reads the entire conversation history with no cache hits,
00:08:40.705 → 00:08:42.485
even though the context is identical.
00:08:42.485 → 00:08:44.705
The Opus plan model setting resolves to Opus
00:08:44.705 → 00:08:46.585
during plan mode and Sonnet during execution.
00:08:46.585 → 00:08:48.645
So each plan toggle is a model switch
00:08:48.645 → 00:08:50.305
and starts a fresh cache.
00:08:50.305 → 00:08:52.845
So it's very interesting because typically the point of that
00:08:52.845 → 00:08:54.025
is to save your session limit.
00:08:54.025 → 00:08:55.985
and I think ultimately in the long run it does,
00:08:55.985 → 00:08:57.105
but it is important to understand
00:08:57.105 → 00:08:59.265
that doing that does reset the cache.
00:08:59.265 → 00:09:02.745
Now what you can do is you can edit your Cloud.MD
00:09:02.745 → 00:09:03.745
and you can do that mid-session
00:09:03.745 → 00:09:05.465
because the edit actually doesn't apply
00:09:05.465 → 00:09:08.825
until you restart that session, so the cache stays safe.
00:09:08.825 → 00:09:11.105
And then of course the Cloud.AI projects caching,
00:09:11.105 → 00:09:12.345
it's not exactly documented,
00:09:12.345 → 00:09:15.145
but pretty confident that it does help to drop docs
00:09:15.145 → 00:09:16.705
in projects rather than in the chat.
00:09:16.705 → 00:09:18.785
But anyways, this token dashboard, like I said,
00:09:18.785 → 00:09:21.945
is very helpful to just be able to understand,
00:09:21.945 → 00:09:24.025
get a little bit more visibility into your tokens.
00:09:24.025 → 00:09:27.465
这个组织的组织的组织
00:09:27.465 → 00:09:29.465
如果你转过去的电话
00:09:29.465 → 00:09:31.225
你的电话会会更改变
00:09:31.225 → 00:09:33.625
比较多于你的主要是用的
00:09:33.625 → 00:09:34.585
很简单的组织
00:09:34.585 → 00:09:35.665
它是一组织的
00:09:35.665 → 00:09:36.985
你会在我的社会里面的
00:09:36.985 → 00:09:37.865
在这里的组织
00:09:38.825 → 00:09:40.665
你会在教授的组织
00:09:40.665 → 00:09:42.425
然后你会在那里的组织
00:09:42.425 → 00:09:43.625
然后你会在那里的组织
00:09:43.625 → 00:09:44.265
你会在这些织织的织织
00:09:44.265 → 00:09:45.465
就要给你一个组织的织织
00:09:45.465 → 00:09:47.465
说说这个是一织织的织织
00:09:47.465 → 00:09:48.985
就像这个织织织的织织
00:09:48.985 → 00:09:50.425
就像它们的织织织
00:09:50.425 → 00:09:52.705
and it will pull in all of your past sessions.
00:09:52.805 → 00:09:54.405
So it's not like you're gonna start fresh
00:09:54.405 → 00:09:57.125
as soon as you link in this repo,
00:09:57.265 → 00:09:58.885
it will read your past files
00:09:58.885 → 00:09:59.985
and it will pull in your tokens.
00:10:00.305 → 00:10:01.585
And then of course, I will also include
00:10:01.585 → 00:10:02.645
that session handoff skill
00:10:02.645 → 00:10:03.985
that I just mentioned to you guys.
00:10:04.045 → 00:10:05.125
So I know this one was super quick.
00:10:05.505 → 00:10:06.485
Hopefully this one was helpful though.
00:10:06.785 → 00:10:07.925
It's just important.
00:10:08.105 → 00:10:10.685
Like I said, when I hear about stuff like this,
00:10:10.685 → 00:10:12.225
I love to understand it to the point
00:10:12.225 → 00:10:13.225
where I know how to use it
00:10:13.225 → 00:10:14.625
and I know what's going on under the hood.
00:10:14.965 → 00:10:17.565
But truthfully, if I looked at some of these other articles,
00:10:17.785 → 00:10:19.545
like how in-depth they go
00:10:19.545 → 00:10:21.385
and how much nuance there is.
00:10:21.625 → 00:10:22.725
Most of the stuff right now,
00:10:22.805 → 00:10:23.985
I just don't need to know
00:10:23.985 → 00:10:26.585
because I'm not using the API in this way super heavily.
00:10:26.765 → 00:10:28.165
So the reason I wanted to throw that out there
00:10:28.165 → 00:10:29.665
is because it's important to stay updated
00:10:29.665 → 00:10:30.605
and follow things,
00:10:30.845 → 00:10:31.705
but just understand
00:10:31.705 → 00:10:34.465
what do you really need to know at its core.
00:10:34.705 → 00:10:35.585
So if you guys enjoyed the video
00:10:35.585 → 00:10:36.265
or you learned something new,
00:10:36.505 → 00:10:37.025
please give it a like,
00:10:37.065 → 00:10:37.705
it helps me out a ton.
00:10:38.545 → 00:10:40.205
I appreciate you guys making it to the end of the video
00:10:40.205 → 00:10:41.405
and I'll see you on the next one.