Pratique du Shadowing: Claude Code Is Wasting Tokens... Microsoft Built a Fix - Apprendre l'anglais à l'oral avec la vidéo

Choisissez un mode d'entraînement

Chaque mode ouvre une salle d'entraînement en plein écran.

119 phrases

Si les phrases sont trop courtes ou trop longues, cliquez sur Edit pour les ajuster.

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.

Télécharger l'application

Tout pour parler couramment

Prononciation IAChaque phrase notée

Phonétique (IPA)Maîtrisez chaque son

VocabulaireEnrichissez votre lexique

Jeu de vocabulaireApprenez en jouant

Scanner le QR

Download on theApp Store Get it onGoogle Play

FOR YOU

Vidéo recommandée

223 phrases

9 Insane Claude Mods I Already Can't live Without

Daily ConversationUniversity StudentsWorking Adults

407 phrases

Why AI Tokens are so Expensive - Computerphile

Daily ConversationUniversity StudentsWorking Adults

207 phrases

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

Daily ConversationUniversity StudentsWorking Adults

78 phrases

Vocabulaire et conseils d’expression pour cette leçon

Cette leçon d’expression orale s’appuie sur la vidéo « Claude Code Is Wasting Tokens... Microsoft Built a Fix ». Les mots qui reviennent le plus souvent : token, agent, economizer, tool, repo. Cette vidéo contient 119 phrases et 1344 mots à répéter en shadowing. La partie parlée dure 7:29. Le locuteur parle vite, environ 180 mots par minute : attendez-vous à des liaisons et à des sons réduits. 82 % des mots font partie des 3 000 mots les plus courants en anglais ; le reste mérite d’être vu avant de commencer.

Vocabulaire clé de cette vidéo

15 mots de la vidéo à retenir, avec leur prononciation et leur sens :

Mot	Prononciation	Sens
token nom	/ˈtoʊkən/	symbole
install verbe	/ɪnˈstɔl/	installer
clone nom	/kloʊn/	clone
unused adjectif	/ʌnˈjuzd/	inutilisé
cloud nom	/ˈklaʊ̯d/	nuage
skill nom	/skɪl/	habileté, compétence
loop nom	/luːp/	boucle
recover verbe	/ɹɪˈkʌvɚ/	recouvrer
subscribe verbe	/səbˈskɹaɪb/	abonner, s'abonner
catalogue nom	/ˈkæt.əˌlɔɡ/	catalogue
prompt verbe	/pɹɑmpt/	inciter
straightforward adjectif	/ˌstɹeɪtˈfɔː(ɹ)wə(ɹ)d/	facile, aisé
folder nom	/ˈfoʊldɚ/	dossier, chemise
refresh verbe	/ɹɪˈfɹɛʃ/	revigorer, rafraîchir
trick nom	/ˈtɹɪk/	tour

Les verbes à particule que vous entendrez

Mot	Prononciation	Sens
come out verbe	/ˌkʌm ˈaʊt/	sortir
turn off verbe		éteindre

Des phrases à répéter

Des phrases courtes et complètes de la vidéo, à réutiliser dans la conversation de tous les jours :

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.

La grammaire de cette vidéo

Les structures que le locuteur utilise le plus, avec les mots exacts de la vidéo :

Structure	Dans la vidéo
Present perfect have/has + participe passé — une action passée qui compte encore maintenant	have been put · you've probably seen · has learned
Voix passive be + participe passé — l’accent est mis sur ce qui arrive, pas sur qui le fait	have been put · are turned · is called

Prononciation à surveiller

Le locuteur utilise 35 contractions et formes réduites, comme I'm, you're, I'll. Prononcez-les sous leur forme courte, telles que vous les entendez.

Les sons « sh » et « zh »: combinations /kɑmbɪˈneɪʃənz/, refresh /ɹɪˈfɹɛʃ/, extension /ɪkˈstɛnʃən/, intervention /ɪntɚˈvɛnʃən/, instruction /ɪnˈstɹʌkʃən/
Mots longs — placez bien l’accent: combinations /kɑmbɪˈneɪʃənz/, intervention /ɪntɚˈvɛnʃən/, consistently /kənˈsɪstəntli/, eliminate /ɪˈlɪm.ɪ.neɪt/, repository /ɹɪˈpɒz.ɪ.t(ə.)ɹɪ/

Les sons difficiles pour les francophones :

/tʃ/ et /dʒ/ — à ne pas adoucir en « ch » et « j »: bench /ˈbɛnt͡ʃ/, usage /ˈjuːsɪd͡ʒ/
/r/ anglais — langue recourbée, sans frotter la gorge: recover /ɹɪˈkʌvɚ/, micro /ˈmaɪ.kɹoʊ/, subscribe /səbˈskɹaɪb/, prompt /pɹɑmpt/, straightforward /ˌstɹeɪtˈfɔː(ɹ)wə(ɹ)d/

Comment s’entraîner avec cette vidéo

Écoutez la vidéo en entier une fois sans parler et notez les mots que vous ne connaissez pas.
Commencez à la vitesse 0,75×, répétez phrase par phrase, puis revenez à la vitesse normale quand cela devient facile.
Enregistrez-vous et comparez avec l’original, en faisant attention à des mots comme token, install, clone.

Qu'est-ce que la technique du Shadowing ?

Le Shadowing est une technique d'apprentissage des langues fondée sur la science, développée à l'origine pour la formation des interprètes professionnels. Le principe est simple mais puissant : vous écoutez de l'anglais natif et le répétez immédiatement à voix haute — comme une ombre suivant le locuteur avec un décalage de 1 à 2 secondes. Les recherches montrent une amélioration significative de la précision de la prononciation, de l'intonation, du rythme, des liaisons, de la compréhension orale et de la fluidité.

Technique du shadowing : lire le guide complet étape par étape →

Pratique du Shadowing: Claude Code Is Wasting Tokens... Microsoft Built a Fix - Apprendre l'anglais à l'oral avec la vidéo

Choisissez un mode d'entraînement

Télécharger l'application

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

Vocabulaire et conseils d’expression pour cette leçon

Vocabulaire clé de cette vidéo

Les verbes à particule que vous entendrez

Des phrases à répéter

La grammaire de cette vidéo

Prononciation à surveiller

Comment s’entraîner avec cette vidéo

Qu'est-ce que la technique du Shadowing ?