쉐도잉 연습: Ultimate Claude Guide: How to Use Claude AI for Beginners in 2026 - 영상으로 영어 말하기 배우기

연습 모드 선택

각 모드는 전체 화면 연습실로 열립니다.

392 문장

문장이 너무 짧거나 길면 Edit를 눌러 조정하세요.

If you're still picking between three Claude models, that system doesn't exist anymore.

Anthropic quietly reshuffled the entire lineup and added a tier above the flagship model.

Most creators haven't caught up, today you do.

You sign up at Claude.ai with Google, Apple, or an email, and there's no credit card at the door.

Free gets you a real working Claude on the mid-tier model, capped by a rolling five-hour message window, and you get up to five projects.

You don't get Claude code and you don't get co-work on free.

Enterprise is custom pricing, sold through sales, and it adds SSO, audit logs, data residency, and compliance controls.

If you need that, you already have a procurement process for it.

At the top of any conversation, the model selector sits right next to the send button.

Click it and you pick which model you're talking to.

And right next to that is the effort and thinking control.

How hard Claude works on each response.

We'll get into what that actually does in a few minutes, because it changed completely in this generation.

Custom Instructures The one setting that changes every answer.

Here's what I actually have in mind.

I run a YouTube channel covering AI tools.

Default to direct specific answers.

Skip the hedging and the four option lists.

When you give me a number, tell me where it came from.

Push back if you think I'm wrong instead of agreeing by default.

That's it.

sentences and it changes the shape of every single response Claude gives me from that point forward.

The mistake most people make is writing something vague like be helpful and concise.

That doesn't constrain anything.

Write it like you're onboarding a new hire.

What to do, what to never do, and what tone to use.

This one setting takes 90 seconds to write

and it's probably the single highest leverage five minutes you'll spend in this entire video.

Projects solve a specific problem.

You stop re-explaining yourself every time you open a new chat.

Each project is its own workspace with its own knowledge base, its own instructions, and on paid plans, its own retrieval system that scales past the normal context limit.

I'm creating one here for a client research stream.

I upload the brand guide, the last two months of reports, and a competitor teardown.

From this point forward, every chat inside this project already has that context.

I never paste those documents again.

On the free plan, you get up to five projects total.

Paid plans give you unlimited projects, and when a project's knowledge base gets big enough to bump against the context limit,

Claude automatically switches into retrieval mode that expands capacity roughly tenfold without you doing anything.

If you're on a team or enterprise plan, you can share a project with specific people or your whole organization.

Two permission levels Can use means they can chat inside the project but can't touch the knowledge base or instructions.

Can edit means they can add documents, update instructions, and shape the project alongside you.

That's the difference between everyone on your team prompting from a blank slate and everyone starting from the same specialist context.

Memory, the new chat that already knows you.

Memory is different from projects.

Projects hold documents for one specific stream of work.

Memory holds facts about you and your patterns across everything you do outside of projects.

Turn it on in settings under capabilities.

Once it's running, Claude automatically summarizes your conversations roughly every 24 hours and builds a running picture.

Your role, your recurring projects, your work and style.

You don't have to ask it to remember anything.

But you can!

Just tell Claude directly what you want kept, and it updates the summary immediately instead of waiting for the next cycle.

Here's the demo I actually want you to see.

Two weeks ago, I told Claude in a normal chat

that I train four times a week on a push-pull split and that my left shoulder gets cranky on overhead press.

I never mentioned it again.

That's the whole feature in one screenshot.

New conversation, zero context pasted in, and it already knows.

Each project has its own separate memory space too.

So what Claude remembers about your client work in one project never bleeds into a personal chat outside of it.

There's also a chat search feature that rides along with memory.

You can search your entire conversation history the way you'd search email, or just ask Claude directly, what were we working on last Tuesday?

And it pulls up the thread.

One housekeeping habit worth building, open Settings Memory once a month and delete anything stale.

Claude is happy to remember things forever, even after they stop being true.

And if you're about to discuss something sensitive you don't want stored at all, there's an incognito toggle, top right of any standalone chat,

that skips memory entirely for that conversation.

This is the one most people misunderstand, because the old mental model is dead and nobody sent out a memo.

Under the old system, extended thinking used to be a manual toggle.

You turned it on, and then you

or whoever built the integration had to set a fixed token budget for how much Claude was allowed to think before answering.

You are basically guessing every single time.

If the budget was too small, Claude cut its reasoning short on a hard problem.

If the budget was too large, you paid for thinking you did not actually need.

The new system is called Adaptive Thinking, and it is the default now on Sonnet 5 and above.

Instead of you setting a fixed budget, Claude looks at the actual complexity of what you are asking and decides for itself

whether to think at all and how hard.

You control the ceiling with an effort level, not the exact token count.

There are four effort levels available to you, and they run from low up to max.

Let me show you the difference directly on the same prompt.

Plan a content calendar for a YouTube channel about AI tools.

Include video topics, post and schedule, and promotion strategy for the next four weeks.

Same prompt run three times.

Low effort, a quick list of four topics, one-line descriptions, no reasoning, response in seconds, medium effort, topics with rationale,

posting times with audience logic explained, a basic promotion framework.

And Matt's effort takes this one step further in a way that genuinely surprised me.

I asked it to help me plan a content calendar for my channel, and instead of just spitting out a plan, it stopped and asked me three clarifying questions first.

It wanted to know what niche the channel actually covers, roughly how many subscribers I'm working with and how often I ship videos right now.

None of that was in my prompt and honestly, none of it would have occurred to me to include upfront.

But the plan it produced after those three answers was miles more useful than what low

or high effort gave me on the same task.

That's the real shift at max.

It stops guessing about your context and starts interviewing you like a decent consultant would.

Low is your default, for anything routine.

Use it on simple questions and quick edits, and don't waste the ceiling on those.

High is where you land for real planning work or anything with strategic stakes.

X-High is what you reach for on genuinely complex problems, where you want Claude to think through every angle.

Max is for the rare case, where you want the most thorough answer possible and you are willing to wait for it.

One thing worth knowing before we move on.

100

On the Mythos tier models, Fable 5 and Mythos 5, thinking can't be turned off at all.

101

It's always adaptive on every single response.

102

That's part of what makes them behave less like a chatbot and more like an agent that reasons before it acts,

103

which matters a lot once we get to the Fable 5 section later in this video.

104

All this Claude 5 work pushed me to finally update my LinkedIn.

105

I had a B2B pitch landing Monday.

106

Claude nailed the bio in 10 minutes, but then I scrawled up to my profile photo and froze.

107

Blurry conference selfie from 2023.

108

That kind of shot quietly kills a pitch, no matter how tight the copy is.

109

So for this launch, I finally went and used InstaHeadshots to fix the visual half of the problem.

110

I'm going to show you this live right now.

111

I already uploaded my shots and picked a style and the model trains on your face for under 15 minutes.

112

The best part is generating the pack is completely free.

113

You only pay if you actually like what comes back.

114

The pack just came back.

115

Look at this one right here.

116

That is genuinely me.

117

My wife asked when I snuck off to a photographer.

118

I immediately ran it for the whole team too, and their reaction was exactly the same.

119

And I'm clearly not alone.

120

Over 36 million headshots generated, 421,000 users, 4.9 stars on Google,

121

and a perfect 5 out of 5 on Trustpilot.

122

Teams at Microsoft, Google, Goldman Sachs, and Harvard, I'll use this.

123

Claude handles the words, but the first thing anyone actually sees is your face.

124

Link is in the description, and there's a discount code right next to it for your first pack.

125

Artifacts are the reason Claude stopped feeling like a text box and started feeling like a workspace.

126

Anytime Claude generates something substantial and self-contained, code, a document, a diagram,

127

a full webpage, it opens in a dedicated panel next to the conversation instead of scrolling past in the chat.

128

Build a simple monthly budget calculator as a single interactive page.

129

Income at the top, expense categories below, a running total, and a bar showing how much is left over.

130

That renders live right there, and I can actually use it, not just read about it.

131

Artifacts aren't limited to code.

132

Turn on code execution and file creation in settings, and Claude can hand you real downloadable files.

133

Word documents, Excel spreadsheets, PowerPoint decks, PDFs built with actual formulas and formatting,

134

not just plain text pasted into a document shape.

135

Every artifact you build lives in the dedicated artifacts tab in your sidebar so it doesn't get buried in chat history.

136

Click share and you get a shareable link.

137

The other person doesn't even need a Claude account to open and use what you build.

138

And paid plans get persistent storage, so a tool you build today still has its data next time you open it.

139

Two apps, one sitting browser tab, no external dependencies, and both shipped in under 10 minutes.

140

That workflow is exactly what we run inside AI Master every day to produce this channel.

141

The entry point is simple.

142

Every top AI model in one window.

143

Instead of paying separate subscriptions to Anthropic, OpenAI, and Google, everything runs from one place.

144

And because of how the platform handles tokens, you pay less than going direct on each model separately.

145

But that's just the starting point.

146

AI Master also handles image generation right in the same interface, and it handles audio and video generation too under one shared token economy.

147

If you build content around a character or persona, there's a consistent character system where your character looks identical from generation to generation.

148

You can publish that character directly on the platform and monetize it.

149

You can also share any generation with the community in one click.

150

Speaking of community, AI Master has 12,000 plus users right now.

151

These aren't lurkers.

152

It's an active network of people shipping real projects with this pipeline every day.

153

There's a seven-day money-back guarantee.

154

If it doesn't click with your workflow, you get a full refund.

155

No questions asked.

156

Head to the link in the description right now.

157

If you buy today, the landing page automatically applies a discount.

158

No code needed.

159

It's time sensitive.

160

So the offer is live while you're watching.

161

That's the whole pitch.

162

Back to the models.

163

Claude has two ways to pull in information beyond its training data and using the wrong one wastes your time either way.

164

Web search is for fast, factual, single answer questions.

165

What's the weather?

166

What did a company just announced?

167

Who won last night's game?

168

It runs one or two lookups and comes back in seconds.

169

Research.

170

What used to be called deep research is a completely different animal.

171

It's a gentic.

172

Claude plans a research strategy, runs five or more searches that build on each other,

173

reads across sources, and comes back with a structured, cited report.

174

This takes anywhere from a couple of minutes to significantly longer, depending on how deep the question goes.

175

That table would have eaten an afternoon of tab juggling if I did it manually.

176

The rule I use, web search when I need one fact fast, research when I need a structured answer I can act on

177

and I'm willing to wait a few minutes for it to actually be right.

178

Okay, that is the foundation for you.

179

You've now got registration handled, the interface mapped, your custom instructions written, project setup, memory turned on,

180

adaptive thinking understood, Artifacts Working, and Research on Tap.

181

Those pieces are the shared workspace every model plugs into.

182

Now let's get to the reason you clicked on this video, which is the five models and exactly which one to run on what.

183

Before we dive into any single model, let me show you the shape of the lineup because that's the piece nobody explains clearly.

184

Anthropic ships five models under the CLAWD5 banner right now, and they are stacked in a ladder from cheapest and fastest at the bottom

185

to most capable and most restricted at the top.

186

Haiku 4.5 sits at the bottom.

187

As the cheap, fast workhorse, Sonnet 5 sits in the middle.

188

And this is where the biggest surprise of the generation lives, because Sonnet 5 is beating the current Opus flagship on knowledge work benchmarks.

189

Opus 4.8 sits above Sonnet, still the frontier for a hard multi-file code and Olympiad-level math.

190

Above Opus, there's a brand new tier that didn't exist in the previous generation, called the Mythos class.

191

Fable 5 lives in that tier as the autonomous project model, and Mythos 5 sits at the very top, with access gated behind a program called Project Glasswing.

192

Let's start with Sonnet 5, because this is the model you should be running on most tasks right now.

193

I opened a blank folder and gave it one prompt with no starter code and no template.

194

I wanted to see if it could ship a real language learning app from scratch in a single shot.

195

Build a gamified Spanish learning app as a single HTML file with no external dependencies.

196

The app should feel like a mini Duolingo.

197

Include a home screen with an XP counter, daily streak flame, and a heart system that regenerates over time.

198

Add three playable mini games, a word to image match with four picture tiles, a sentence builder puzzle where the user drags Spanish word tiles into the correct order,

199

and a listen and type game that reads a Spanish word out loud using the browser speech API.

200

Add a spaced repetition cue so missed words come back sooner.

201

Use the Claude API with a pleaseholder key to generate example sentences on demand.

202

Persist all progress to laestertage.

203

Add level up animations with Confetti when the user hits a new XP threshold.

204

Sonnet 5 shipped the file in about 7 minutes and it worked on the first open.

205

The home screen rendered with a glowing streak flame and a bouncing XP bar at the top.

206

I clicked into the word match game and got four image tiles with a Spanish prompt above them.

207

The sentence builder puzzle let me drag tiles into slots with a satisfying snap animation.

208

The listen and type game actually spoke the Spanish word out loud through the browser's speech synthesis.

209

The part that really impressed me was not the visuals.

210

Sonnet 5 wired up a proper spaced repetition queue without being asked to explain the algorithm.

211

Words I missed came back three cars later, and the app queued them again after 10 more turns.

212

The confetti animation on my first level up ran smoothly at 60 frames a second.

213

The hard regeneration timer even persisted across page reloads.

214

It felt like a shipped product, not a first-pass prototype.

215

A year ago, you needed Opus to get that kind of output.

216

Now you get it at Sonnet pricing, which during the intro window runs $2 per million input tokens.

217

That window closes at the end of August 2026, and then it goes to $3 per million.

218

Still a fraction of what Opus costs, and you are getting output that Sonnet had no business producing 12 months ago.

219

Now our second scenario, this time, pure browser.

220

No server, no dependencies whatsoever.

221

I wanted something visual enough that you could hand it to a non-technical person and they would immediately see what it does.

222

So I asked Sonnet Phi to build an interactive knowledge graph.

223

Notes you can drag, edges that spring between them, a physics simulation running entirely on canvas.

224

Build an interactive network graph visualizer as a single HTML file with no external libraries.

225

The user types a topic into a search box and the graph auto-generates related concept nodes around it.

226

Nodes should be draggable with a spring physics simulation on Canvas.

227

Click in a node, expand it into its own cluster.

228

Include a sidebar that lists all nodes in the current graph with a delete button for each.

229

Color code clusters automatically.

230

Everything persists in LeisterPage, so the graph is still there on the next page reload.

231

Ship it as one file, under 300 lines.

232

It came back in about 8 minutes.

233

I typed machine learning into the search box

234

and watched about 14 nodes pop out from the center with physics animated edges snapping into place.

235

I dragged the node across the screen and the connected nodes followed it on springs.

236

I clicked neural networks and it expanded into its own subcluster.

237

The sidebar listed every node with a live delete button.

238

The whole thing runs at 60 frames a second and reloads exactly where you left it.

239

The detail that got me was the physics engine.

240

Sonnet 5 implemented its own spring simulation from scratch in under 40 lines of vanilla JavaScript.

241

No library, no shortcut.

242

It tuned the damping coefficient so nodes settle instead of oscillating forever.

243

I did not specify any of that.

244

It just reasoned that a graph with nodes bouncing indefinitely would feel broken, and it fixed that before I could ask.

245

That is the shift that makes Sonnet 5 a genuine default and not just a cheaper substitute.

246

It is not just producing code, it is reasoning about the whole context of how someone would actually use what it is building,

247

and then filling in the gaps you did not think to specify.

248

There is one caveat that Anthropic buried, and you need to know about it before it bites your automations.

249

Sonnet 5 uses a new tokenizer.

250

The same text you were sending to older Claude models now generates roughly 30% more tokens.

251

Your old max underscore token settings are wrong, your old cost estimates are wrong, and if you have any automations firing against Sonnet 5 today,

252

you need to recalibrate those before the intro pricing window closes in August.

253

Opus 4.8 sits one tier above Sonnet in the lineup.

254

It costs $5 per million input tokens and $25 per million output.

255

The tasks where it earns that price are the ones that require holding a large, complex structure in memory all at once.

256

Deep document analysis, long-form knowledge work, anything where every piece has to stay coherent from start to finish.

257

Here is the first task I gave it.

258

I wanted Opus to build an entire online course from scratch.

259

Not just a syllabus, but a complete learning experience with modules, exercises, quizzes, and a difficulty curve that actually makes sense for a beginner.

260

The topic.

261

SQL for product designers who've never touched a database in their life.

262

That is a hard brief because you have to hold the student psychology in your head the whole time, not just the technical facts.

263

Here is the prompt I gave it.

264

You are a curriculum designer.

265

Build a complete beginner SQL course for product designers with zero database experience.

266

Structure it into five modules.

267

Each module needs a clear learning goal, three concept explanations with real product design analogies,

268

two hands-on exercises with sample data, and a five-question quiz.

269

And with a capstone project brief where the student designs a data model for a real app feature.

270

Adjust the difficulty so a complete beginner finishes module one feeling confident, not overwhelmed.

271

Opus held the pedagogical logic across the entire course, and the analogies it chose felt like they were written by someone who actually knew both worlds.

272

That coherence across eight interconnected sections is exactly where Opus earns its price.

273

The second use case is financial document analysis.

274

I took Apple's most recent 10k annual report, uploaded it to Opus and asked for a full investment memorandum.

275

Not a summary, a memo with a real opinion, specific risk flags and numbers pulled from the document.

276

That contradiction catch is the thing that matters.

277

Opus read across two different sections of a 200-page document and noticed that the numbers did not reconcile.

278

It went back, cross-checked, and surfaced the conflict without being asked.

279

That is the deep context gap.

280

When your task requires holding a large, complex document together as a single object and finding the places where it contradicts itself,

281

Opus is the model for that.

282

Opus is not worth it on a bunch of tasks.

283

Skip it on long context summarization and on plain data extraction.

284

Skip it on standard content generation too.

285

Skip it on any code refactor under a thousand lines.

286

Save it for tasks that genuinely require holding a large complex structure in memory all at once

287

and reasoning about how every piece connects to every other piece.

288

Now, if Opus is the escalation model for deep single session work, Fable 5 is a completely different animal.

289

It is not for a task, it is for a project that runs while you sleep.

290

Fable 5 sits in a brand new tier that did not exist in the previous generation.

291

Anthropi calls it the Mythos class, and it lives one full step above Opus.

292

Sonnet and Opus are models you prompt.

293

Fable 5 is a model you brief.

294

Let me show you what this looks like on a smaller, more realistic project.

295

I gave Fable 5 the following brief, kicked it off at 11 at night, and went to sleep.

296

Objective tokens, NAM build a working Chrome extension

297

that summarizes any YouTube video into a five bullet summary and a copyable clipboard block.

298

Constraints, MV3 only, no external servers.

299

Use the Cloud API with a policeholder key.

300

Budget, eight hours of wall time, unlimited tool use.

301

Deliverables, complete extension folder.

302

Read me with install steps.

303

One test video demonstrating the extension working.

304

If you hit an unrecoverable error, log it and continue with the next module.

305

I woke up at 7 the next morning.

306

The extension was in the folder.

307

The readme was written.

308

There was a test recording showing it working on a real YouTube video.

309

There were also three log entries where Fable had hit errors, diagnosed them, and rerouted.

310

It even flagged one design decision it was not sure about and left a note asking me to confirm.

311

The thing that makes Fable 5 different is not raw intelligence, it is durability.

312

It does not fall over when something breaks.

313

It plans around failure.

314

That is the capability jump the Mythos class is pointing at.

315

One flag before you get too excited.

316

The Fable 5 benchmark numbers Anthropic is publishing are from their own system cards and a couple of third-party trackers.

317

everything has been independently reproduced yet.

318

Treat those benchmarks as anthropics numbers, not as ground truth.

319

The Stripe story, on the other hand, is a first-party client quote, and you can trust that one.

320

Pricing on Fable 5 is $10 per million input tokens and $50 per million output.

321

That is real money, but the calculation is not cost per token, it is cost per finished project.

322

An 8-hour autonomous Fable run

323

that ships a working artifact is often cheaper than 3 days of your own time driving Opus manually.

324

Fable delegates most of its execution work to Haiku 4.5.

325

I had a folder with 200 product entries in a CSF.

326

Each row needed a unique, channel-appropriate description written from scratch in under 30 words.

327

I gave Claude a single instruction.

328

Process this CSV, spin up as many Haiku sub-agents as you need, and merge the outputs back into a clean file.

329

Here is what happened.

330

Haiku 4.5 is also the model I recommend for high-volume operational work.

331

Think about customer support triage, where you route thousands of tickets an hour.

332

Log summarization on production systems is another obvious fit for it.

333

Email classification on a busy inbox works cleanly too.

334

And content moderation running at real scale is where Haiku really shines.

335

Anywhere you need thousands of calls per hour, Haiku is the answer.

336

And that brings us to the next model, which is honestly the strangest one in the whole lineup.

337

Mythos 5 sits at the top of the ladder, above Fable in the same Mythos tier.

338

But you cannot just log in and use it today.

339

Access runs through a program Anthropic calls Project Glasswing,

340

and right now that means a curated list of research labs and enterprise partners who applied and got selected.

341

What Anthropix says publicly is that Mythos 5 capabilities exceed anything they have ever made generally available.

342

That puts it above Fable 5, which is already a significant step up from Opus.

343

In practice, Mythos can hold longer planning horizons and tackle harder reasoning chains than Fable can sustain.

344

The organizations with access are using Mythos 5 on problems that a human team would take months to solve.

345

We are talking about long horizon scientific research and enterprise workflows that span dozens of interconnected systems over several weeks.

346

They are watching what happens when a much more capable model gets much more autonomy.

347

They want that playing out in monitored environments before it reaches everyone.

348

Here is my honest read on this.

349

I have not had access to Mythos 5 myself, and most of you watching this have not either.

350

But the line Anthropic drew here is itself a signal.

351

When Glasswing eventually opens up more broadly, the use cases are going to look very different from anything in this video.

352

Before we wrap up, there are three brand new capabilities that landed with this whole generation, and any of the five models we just covered can use them.

353

These live in the workflow layer on top of the models, not inside a single model, so they matter no matter which one you end up running.

354

Let's walk through the three that actually change how the day looks.

355

Catwork is the feature that finally makes Claude feel like a team, not just a chat window.

356

You give it one project, and it runs multiple agent threads in parallel on that same project.

357

Each thread is a full conversation with its own agent, and all of them share the same context underneath.

358

Let me show you what that looked like on a video project I actually ran this morning before recording this.

359

That whole run wrapped up in about 11 minutes while I was still on my first coffee.

360

In the old world, I'd have opened three separate chats and wasted 20 minutes re-explaining the same project context each time.

361

Cowork just holds that context once and lets three agents work off it in parallel.

362

Now let me pick up one artifact from that co-work run and show you the next feature.

363

The thumbnail concepts came back as rough written descriptions, so I dropped one of them into Claude Design to see what the actual visual would look like.

364

Design shifts baked into every Claude 5 model, so you skipped the Figma round trip entirely.

365

So the loop I just ran goes from a rough co-work brief to a real style thumbnail concept without ever leaving Claude.

366

I still hand the final version to my designer for polish.

367

But that concept phase that used to eat half a day now takes about two minutes end to end.

368

The last piece of the workflow layer lives inside Claude Code, and Anthropic calls it dynamic workflows.

369

This is where a bigger model like Sonnet or Opus spawns hundreds of parallel haiku sub-agents on one task.

370

I showed you a small version of this earlier with the 200-row CSV.

371

The same pattern scales all the way up to thousands of files at once.

372

Now that you've seen what each model actually does, the plan question gets simple.

373

You already know which capabilities you need, so let's map them to tiers in about a minute.

374

Free gives you a working clod on the mid-tier model with a rolling 5-hour message window, and it's enough to follow along with most of this video.

375

You don't get clod code and you don't get co-work on Free, and Opus plus Mythos tier access, stay locked behind a paid plan.

376

Pro sits at 20 bucks a month, or 17 if you pay annually, and this is the plan almost everyone watching should be on.

377

It unlocks Clot Code, Cowork, Research, Adaptive Thinking, and the full model lineup including Mythos tier access.

378

Max runs 100 a month for 5 times Pro's usage, or 200 for 20 times.

379

It's the same feature set with a bigger usage bucket, so only jump to max if you're actually hitting pros limits in your daily work.

380

Team and enterprise exist if you need shared projects or centralized billing, and you already know if that's you.

381

Max runs 100 a month for 5 times pros usage, or 200 for 20 times.

382

It's the same feature set with a bigger usage bucket, so only jump to max if you're actually hitting pros limits.

383

Free gets you a work in Claude on the mid-tier model with a rolling 5-hour message window.

384

You don't get Claude code, and you don't get co-work on Free.

385

Opus and Mythos-tier access also stay locked behind a pay plan.

386

That is the full Claudia 5 generation for you.

387

You've got 5 models and one clear decision framework to pick between them.

388

Run Sonnet 5 for everything by default.

389

Eskescalate to Opus when the task gets genuinely hard on you.

390

Reach for Fable when you want to brief an autonomous agent and walk away for hours.

391

Keep hockey running as your executor layer underneath the bigger models.

392

Mythos 5 stays on the horizon for the rest of us until Glasswing opens up.

앱 다운로드

유창하게 말하는 데 필요한 모든 것

AI 발음 평가문장마다 점수 확인

IPA 발음 연습모든 소리 마스터

단어 학습나만의 단어장 만들기

단어 게임놀면서 외우기

QR 스캔

Download on theApp Store Get it onGoogle Play

FOR YOU

실생활에서 유용한 AI 도구 사용법 대화

이 영상은 2026년 클로드 AI를 처음 사용하는 사람들을 위한 가이드로, 실제로 AI 도구를 활용할 때 필요한 가입 방법, 모델 선택, 프로젝트 관리 등을 다룹니다. 영어 회화 연습이나 IELTS 스피킹 준비에 도움이 되는 실용적인 대화 내용이 담겨 있어, 영상 영어 공부를 하면서 실전적인 표현을 익힐 수 있습니다.

외워두면 유용한 표현과 구문

rolling five-hour message window: 5시간 회전형 메시지 창 (시간 제한이 있는 서비스 환경에서 자주 사용되는 표현)
push back: 반박하다, 의견을 제시하다 (상대방의 잘못된 생각을 지적할 때 사용)
high leverage: 높은 효과를 발휘하는 (시간이나 노력에 비해 큰 결과를 가져오는 일을 설명할 때)
retrieval mode: 검색 모드 (AI가 저장된 정보를 찾아 사용하는 기능을 나타냄)
recurring projects: 반복적인 프로젝트 (자주 수행하는 작업을 설명할 때)

지금 바로 할 수 있는 shadow speech 챌린지

영어 발음 교정과 회화 실력을 키우기 위해, 이 영상의 대화를 3번 따라 읽어보세요. 첫 번째는 원래 속도로 듣고 따라하기, 두 번째는 느린 속도로 발음을 확실히 하기, 세 번째는 자신의 말처럼 자연스럽게 말하기입니다. 특히 "push back"이나 "high leverage" 같은 표현을 강조해서 발음하며, 의미를 이해하면서 말하면 더 효과적입니다. 이렇게 shadowing을 하면 실제 회화에서 자연스럽게 표현을 사용할 수 있게 될 거예요!

이 영상의 문법

화자가 가장 많이 쓰는 문형을 영상 속 실제 표현과 함께 정리했습니다.

문형	영상 속 표현
“used to” used to + 동사 — 지금은 아닌 과거의 습관이나 상태	used to be · used to eat
수동태 be + 과거분사 — 누가 하는지보다 무슨 일이 일어나는지에 초점	was allowed · is called · can't be turned
현재완료 have/has + 과거분사 — 과거의 일이 지금도 관련이 있을 때	haven't caught · who've never touched · has been
관계절 who / which + 절 — 사람이나 사물에 대한 추가 정보	video, which is · someone who actually · model, which is

쉐도잉이란? 영어 실력을 빠르게 키우는 과학적 방법

쉐도잉(Shadowing)은 원래 전문 통역사 훈련을 위해 개발된 언어 학습 기법으로, 다언어 학자인 Dr. Alexander Arguelles에 의해 대중화된 방법입니다. 핵심 원리는 간단하지만 매우 강력합니다: 원어민의 영어를 들으면서 1~2초의 짧은 지연으로 즉시 소리 내어 따라 말하는 것——마치 '그림자(shadow)'처럼 화자를 따라가는 것입니다. 문법 공부나 수동적인 청취와 달리, 쉐도잉은 뇌와 입 근육이 동시에 실시간으로 영어를 처리하고 재현하도록 훈련합니다. 연구에 따르면 이 방법은 발음 정확도, 억양, 리듬, 연음, 청취력, 말하기 유창성을 크게 향상시킵니다. IELTS 스피킹 준비와 자연스러운 영어 소통을 원하는 분들에게 특히 효과적입니다.

섀도잉 방법: 단계별 전체 가이드 읽기 →

쉐도잉 연습: Ultimate Claude Guide: How to Use Claude AI for Beginners in 2026 - 영상으로 영어 말하기 배우기

연습 모드 선택

앱 다운로드

추천 영상

How To Use Claude Better Than 99% Of People

I Tried PewDiePie's Odysseus AI So You Don't Have To

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

Can This Robot Be My Personal Assistant

Agentic AI Frameworks Explained: Workflows, Multi-Agent, & Production

A Practical Guide To Becoming An AI Engineer (2026)

When AIs act emotional

Prompt Engineering - Guidelines

New Skills! v1.2 brings /wait-what, /writing-for-agents, and fixes /grill-me

NEW Claude AI Can Create Faceless Youtube Videos In MINUTES!

Top 5 Claude Skills That can save 10x Your AI Sessions

Anthropic just proved AI isn't getting better

실생활에서 유용한 AI 도구 사용법 대화

외워두면 유용한 표현과 구문

지금 바로 할 수 있는 shadow speech 챌린지

이 영상의 문법

쉐도잉이란? 영어 실력을 빠르게 키우는 과학적 방법