跟讀練習: Claude Code Is Wasting Tokens... Microsoft Built a Fix - 通過影片學習英語口語

選擇練習模式

每種模式都會開啟全屏練習室。

119 句

如果句子過短或過長，請點選 Edit 進行調整。

It seems there's a lot of heat with tools and extensions that claim to reduce token usage.

Many of these have been put to the test and they work well.

Now Microsoft just released their own, a free plugin that cut AI coding agent costs by 18 to 31 % in their tests.

And they didn't use a cheaper model or a different agent either.

It works with all the big names, so it's a win for us.

it's called token economizer

and it solves a problem you've probably seen yourself

you ask claude code to fix something fix this it tries a command gets an error

and then keeps running

and burning tokens up to find us a fix token economizer learns from those sessions it remembers how your repo works, removes tools you don't use,

and steps in when the agent keeps making the same mistake.

And the more you use it in the same project, the more tokens you actually save.

So let's install it and see what it actually does.

If you enjoy coding tools the speed of your workflow, be sure to subscribe.

We have videos coming out all the time.

Now I'm going to use

Cloud Code for this, but remember

Token Economizer also supports all the other big names.

First, I'll clone it and install the plugin right here.

We're going to run our git clone.

We're going to install the git plugin from the marketplace and make sure everything is ready to go.

That's installed all super straightforward.

Just keep the clone folder because the plugin runs from there.

Now I'll open one of my existing projects and we're going to fire up Claude Code.

Let's check the status.

All we need to do to check the status here is I'm going to run token economizer status.

This shows what the plugin has learned.

It shows what it's doing and the costs it has already tracked.

Now, quick note here, micro skills are turned off by default.

So I'm going to enable those and we'll circle back.

Then I'll refresh the tool catalog.

I'm going to run here a tool prune, refresh catalog.

Now I'll give Claude Code some tasks.

Let's start super simple.

I'm going to give Claude Code a task, build a unit test.

Very straightforward.

Run it.

Now notice I'm not changing how I use Claude code.

It's the same thing, just on Opus 5 .5, with all the prompts I would normally fire straight at it.

TokenEconomizer just works alongside it.

Now, in a repo where it's already watched several sessions, we can check what it's learned so far.

So here we are.

I'm going to run Microskills.

You can check.

You can run ToolPrune.

Here we can see any repo -specific info it's learned.

In tools, it's identified as unused.

If you install it for the first time, these might be empty.

It needs a few sessions to actually learn what the hell is going on.

And all of that info stays on your machine inside the token economizer folder.

There are really three parts to this.

Now, first up to the plate is called micro skills.

Say you're working on a project that needs a specific command to run its tests.

Claude code, it doesn't know that.

So it tries the usual command, gets an error, try something else and yeah, okay, it eventually gets it.

Great.

Except later when you start a new session, it might make the exact same mistake.

So there we are paying extra tokens, which I would rather not do.

Token Economizer looks at a previous session, all of them, and learns small details about the repo.

Things like which test commands actually work, which environment variables are needed, and where certain files live.

Once something has been confirmed enough times, it can give us that information to the agents at the start of a new session.

So instead of wasting tokens over and over again, the agent already has the answer.

Now next up to the plate is tool pruning.

When you start a coding agent, it can load a bunch of tools into its context, and those tool definitions use tokens,

even if the agent never touches them.

Token Economizer tracks which tools actually get used in your repo.

If a tool consistently goes unused, used, it can stop loading it in a feature session.

And if you need it again, you can just restore it.

Then finally, we go to something called a failure loop intervention.

You've probably watched an agent run a command, get an error, change one small thing and run almost the exact same command again.

Sometimes it just keeps repeating that process.

Token Economizer watches for these patterns.

When it thinks the agent might be stuck, it uses a judge model to check what's happening.

If it detects this loop, it sends the agent a short instruction to try a different approach.

And if the agent still can't recover, it tells you how to stop the turn.

It doesn't just freak out and stop, but does it actually save enough money to even matter?

That's sort of the question.

Yes, it appears to do pretty much that.

Microsoft in UC Berkeley tested

Token Economizer on...

277 SWE benched tasks across six repos using five different agent and model combinations.

And the average cost dropped between 18 .4 % and about 31%.

So the results were pretty good.

In four of the five combinations, tasks' success actually improved.

One saw a small decline, but across all 30 repository and agent model combinations, Token Economizer was cheaper than running without it.

And they even included the plugin's own model calls in those.

Token Economizer isn't completely free to run.

Some of its features use additional model calls.

So those reported savings already accounted for the overhead.

It is open source.

It works with the same agents we already use, and the info it learns stays local.

But I wouldn't expect all of us to see the same results.

Its first public release was in October of this year, and the biggest savings come from working in the same repositories over multiple sessions.

If you're constantly jumping between unrelated projects, there's much less to actually learn from it, or for it, I should say.

If your agent is doing something completely new, there might not be much repeated work to eliminate at all.

So ask yourself, based on the work you do, will this actually be useful?

If you use agents every day on one or a couple core projects, then yeah, this should help.

If you're working on the same repo over and over with Claude Code, Codex, yes, this is going to help.

But if you're building a ton of different projects or you're just playing around, this is not really going to change anything.

100

It's probably not even worth it.

101

So here's what I'd do.

102

Install Token Economizer on the repo you use the most.

103

Run it for a week, then check the status reports like I showed us and compare the costs of similar tasks.

104

If the agent is actually using fewer tokens and whether it's still completing your tasks properly, awesome.

105

It is working.

106

One thing to remember, each agent learns separately.

107

So whatever cloud code learns doesn't automatically carry over to Codex.

108

And don't judge it entirely on the first session.

109

That's why I say run it for a week.

110

The whole point is that it learns from repeated use.

111

If it's saving you money without getting in your way, keep it.

112

If not, you've got the answer.

113

So what it really does is takes a different approach to

114

reducing AI coding costs instead of making the model smaller or compressing your prompts.

115

It looks at the work your agent keeps wasting tokens on.

116

It remembers useful details about your repo, stops loading unused tools, and helps agents recover when it gets stuck.

117

I'm Josh from BetterStack.

118

If you enjoy coding tips and tricks like this, be sure to subscribe.

119

We'll see you in another video.

下載應用

流利口語所需的一切

AI 發音評分逐句打分

音標練習掌握每個音

詞彙學習建立你的詞庫

詞彙遊戲邊玩邊學

掃碼

Download on theApp Store Get it onGoogle Play

FOR YOU

本課的詞彙與口語要點

這節口語課以影片“Claude Code Is Wasting Tokens... Microsoft Built a Fix”為素材。影片中反覆出現的詞有：token, agent, economizer, tool, repo。這段影片共有 119 個句子、1344 個單詞可供跟讀。講話部分時長為 7:29。說話人語速較快，每分鐘約 180 個詞，會出現較多連讀和弱讀。 82% 的單詞屬於英語最常用的 3,000 詞，其餘的詞建議在練習前先學一下。

影片中的重點詞彙

影片中 15 個值得學習的單詞，附發音和釋義：

單詞	發音	釋義
token 名詞	/ˈtoʊkən/	代幣 /代幣, 籌碼 /籌碼
install 動詞	/ɪnˈstɔl/	安裝 /安裝
clone 名詞	/kloʊn/	克隆, 複製 /複製
cloud 名詞	/ˈklaʊ̯d/	йүн, 雲 /雲
skill 名詞	/skɪl/	技巧, 技能
loop 名詞	/luːp/	圈, 環 /環
recover 動詞	/ɹɪˈkʌvɚ/	追回, 收復
subscribe 動詞	/səbˈskɹaɪb/	簽署 /簽署
catalogue 名詞	/ˈkæt.əˌlɔɡ/	目錄 /目錄
straightforward 形容詞	/ˌstɹeɪtˈfɔː(ɹ)wə(ɹ)d/	簡單 /簡單
folder 名詞	/ˈfoʊldɚ/	卷宗, 資料夾 /資料夾
refresh 動詞	/ɹɪˈfɹɛʃ/	提神
trick 名詞	/ˈtɹɪk/	詭計 /詭計, 騙術 /騙術
alongside	/əˌlɔŋˈsaɪd/	旁邊 /旁邊
compare 動詞	/kəmˈpɛɚ/	比, 比較 /比較

影片中出現的短語動詞

單詞	發音	釋義
come out 動詞	/ˌkʌm ˈaʊt/	出, 出來 /出來
turn off 動詞		關 /關, 關掉 /關掉

值得反覆跟讀的句子

影片中簡短而完整的句子，可以直接用在日常對話裡：

We're going to run our git clone.
Then I'll refresh the tool catalog.
That's why I say run it for a week.
If not, you've got the answer.
We'll see you in another video.

影片中的語法

說話人最常用的結構，並附上影片中的原話：

結構	影片中的用法
現在完成時 have/has + 過去分詞 — 過去發生但與現在仍有關聯的事	have been put · you've probably seen · has learned
被動語態 be + 過去分詞 — 強調發生了什麼，而不是誰做的	have been put · are turned · is called

需要注意的發音

說話人用了 35 次縮略和弱讀形式，例如 I'm, you're, I'll。請按聽到的簡短形式來說。

“sh” 和 “zh” 音: combinations /kɑmbɪˈneɪʃənz/, refresh /ɹɪˈfɹɛʃ/, extension /ɪkˈstɛnʃən/, intervention /ɪntɚˈvɛnʃən/, instruction /ɪnˈstɹʌkʃən/
長單詞——注意重音位置: combinations /kɑmbɪˈneɪʃənz/, intervention /ɪntɚˈvɛnʃən/, consistently /kənˈsɪstəntli/, eliminate /ɪˈlɪm.ɪ.neɪt/, repository /ɹɪˈpɒz.ɪ.t(ə.)ɹɪ/

中文母語者容易讀錯的音：

/v/ — 不要讀成 /w/，上齒輕觸下唇: recover /ɹɪˈkʌvɚ/, solve /sɒlv/, intervention /ɪntɚˈvɛnʃən/, variable [ˈvæɹ.i.ə.bl̩], overhead /ˈəʊvəˌhɛd/
詞尾子音 — 要讀清楚，後面不要加母音: combinations /kɑmbɪˈneɪʃənz/, unused /ʌnˈjuzd/, cloud /ˈklaʊ̯d/, loop /luːp/, subscribe /səbˈskɹaɪb/
/r/ 和 /l/ 的區別: variable [ˈvæɹ.i.ə.bl̩], separately /ˈsɛp.ɹət.li/

如何用這段影片練習

先完整聽一遍影片，不要開口，記下不認識的單詞。
先用 0.75 倍速逐句跟讀，熟練之後再回到正常速度。
錄下自己的聲音並與原聲對比，特別注意 token, install, clone 這類單詞。

什麼是跟讀法？

跟讀法 (Shadowing) 是一種有科學依據的語言學習技巧，最初開發用於專業口譯員的培訓，並由多語言者Alexander Arguelles博士普及。這個方法簡單而強大：您在聽英語母語原聲的同時立即大聲重複——就像是一個延遲1-2秒緊跟說話者的影子。與被動聽力或語法練習不同，跟讀法強迫您的大腦和口腔肌肉同時處理並模模擬實的講話模式。研究表明它能顯著提高發音準確性，語調，節奏，連讀，聽力理解和口語流利度——使其成為雅思口語備考和真實英語交流最有效的方法之一。

影子跟讀法: 閱讀完整分步指南 →

跟讀練習: Claude Code Is Wasting Tokens... Microsoft Built a Fix - 通過影片學習英語口語

選擇練習模式

下載應用

推薦影片

9 Insane Claude Mods I Already Can't live Without

Why AI Tokens are so Expensive - Computerphile

Your AI Agent Starts Every Session Like a New Hire - Killian Carlsen-Phelan, Sonar

The easy way to keep your repos tidy.

Did Anthropic Kill Git?

So I tried Matt's skills...

Give Me 24 minutes, I'll Give You 1,000+ hours of Claude Knowledge

I Tested The Viral Reverse Engineering Repo

JUST $1 For Billions of Tokens?! (DeepSeek v4 Pro)

GitHub Copilot now controls your command line...

The 4-Hour Workweek in a World of AI

STOP Running Code On Your Computer