逐字稿文字預覽
Everyone and their mom is talking about one GitHub repo right now. It's called REA, and that stands for Reverse Engineer Anything. You point Claude at an app you don't have the code for, and Claude figures out how it works. Then you can rebuild it from scratch for free. It has over 149,000 stars on GitHub, but here's the weird part. It's six months old, and for almost all of that time, nobody cared. Last Wednesday, it had about Last Wednesday, it had about 13,000 stars. Five days later, it had 149,000. So why is everyone talking about it now? It blew up the same week someone used Claude to build free copies of Photoshop and six other Adobe apps. So people were already asking how far AI can go with copying apps. REA showed up with an answer. One post on X put it like this. We spent years teaching AI to write software. Now we're teaching it to Now we're teaching it to take software apart. When I first saw it, I figured it was a hacker tool I'd never touch, but it actually answers a question I've been stuck on since my last video, which was about those Adobe copies. If anyone can copy an app, what are people going to pay for? So here's a quick test you can run on your own product right now. Ask yourself one question about whatever you sell. If someone copied my product perfectly tomorrow, why? Tomorrow, why would my customers stay? Whatever your answer is, that's your moat. A moat is just the thing that keeps competitors from copying you. And if you can't answer that question, you first need to give your customers a reason to stay. Maybe you answer their messages the same day. Maybe you already know their business inside and out. Maybe they just like working with you. None of those have anything to do with your code. Get one of those first before you worry about anyone copying your app. app. So I'll show you what REA actually does and who should and shouldn't touch it. Then I'll give you the three things people will still pay for when anyone can copy anything. The third one is the one I'm betting my whole business on. If you like building cool things with Claude, I co-founded the Claude Code Club. That's where we build stuff like this together and the links in my description. Before any of that, you need to understand one thing. Every app on your computer is a sealed box. your computer is a sealed box. Photoshop, Notion, the game you played as a kid. You get the box, but you never get the code inside it. For a long time, opening that box took a specialist and a lot of patience. What the REA repo does is hand that sealed box to Claude so Claude can open it up and see how the app works. So what is REA really? Well, REA is a free add on for your agent. It works with Claude Code, Codex Cursor, Gemini CLI and a Gemini CLI, and a few others. It plugs in through something called MCP, which is just a way to give Claude new tools. Installing it is easy. Just paste the GitHub link into Claude code and say, install this for me. Claude sets it up, and then you restart Claude code. Pro tip, before you install it, paste the same link and ask Claude what this repo does and how you could use it. Now, Claude has a toolbox for looking inside apps, so here's how it works under the hood. When a company ships an app, its code gets turned into machine code. That's the ones and zeros your computer runs and a human can't read it. There are other programs that turn machine code back into something a human can read. The big ones are called Ghidra, Hopper and IDA. Ghidra is free and fun fact, the NSA built it. REA lets Claude drive those programs for you. Claude asks a question. REA digs into the app and it shows you the exact spots in the app that prove its answer. app that prove its answer so you know it isn't guessing. figure out. Think of it like a chef at a restaurant. Normal coding with Claude is cooking from a recipe card. You have the formula so you can make it. REA is the chef who can figure out the entire recipe and every ingredient from a single bite. The example prompt in their own repo is simple. You type this into Claude. Understand how search works in the Notes app, show me the evidence, and build a similar feature for my project. feature for my project. Swap notes for any app and swap search for any feature you like. The last part is the line that got everyone talking. Build a similar feature for my project. Another pro tip is give Claude screenshots of the feature you want or tell it to open the app and look for itself. The more Claude can see, the better it does. So how far can it actually go? The use cases people have come up with are wild. So what are people actually building with it? with it? Well, here are three real examples. The first use case comes from a developer named Mukesh Murigan. He pointed REA at the color picker in Power Toys, which is Microsoft's free toolkit for Windows. He only gave Claude the finished app, not the code. But Microsoft's code for that app is public, so he could check REA's answers against what Microsoft actually built. REA got eight out of nine findings fully right. Then he used what it found to build his own color picker in under 750 lines of code. lines of code. The second use case is a game called DX ball. It's a brick breaker game from the 1990s. One developer even used REA to find the exact math the game uses to move the sound from your left speaker to your right as the ball moves. Then they rebuilt that math from scratch. The rebuilt version matched the original game in all 3,205 tests down to the exact bytes. That's a 30 -year -old game rebuilt without ever seeing its code. seeing its code. And the third use case is the one that made me stop scrolling. Another developer pointed REA at a free trial of QuickBooks, the accounting app. Then they built their own open source accounting app with invoices, bank imports, and reports. Even they say REA didn't hand them the original code or a finished product in one step and that it took real work. But think about that
本次執行流程與問題處理
下達指令
預期流程
- 下載與語音辨識
- 來源選擇與品質檢查
- 文字校正與句子單元
- 雙語翻譯
- 雙語時間軸與學習單元
- 筆記與發布
- 公開網頁發布
1. 下載與語音辨識
工具：videonote-transcribe-url（主辨識管線：下載、Whisper 辨識與選稿）
處理：下載音訊、以 Whisper 辨識並選出最佳逐字稿
產出：yt_dlp_meta.json（影片元資料）、audio_16k_mono.wav（16kHz 單聲道音訊）、transcript.whisper.json（Whisper 辨識結果）、transcript.json（正式逐字稿）、transcript_selection_report.json（來源選擇報告）
耗時：953.7 秒
關係：依序執行
結果：完成
2. 建立筆記頁面
工具：videonote-render-html（產生筆記 HTML）
處理：產生可閱讀的筆記 HTML 頁面
產出：note.html（筆記 HTML）
耗時：0.5 秒
關係：依序執行
結果：完成
3. 發布公開頁面
工具：videonote-publish（發布公開頁面）
處理：發布到公開網站與每日列表
產出：公開筆記網址、result.json（管線結果）
耗時：0.7 秒
關係：依序執行
結果：完成
手動來源字幕
自動來源字幕
Whisper 字幕
Qwen 字幕
- 尚未產生
Parakeet 字幕
- 尚未產生
MOSS 字幕
- 尚未產生
聲學時間證據（PyAnnote / WhisperX）
- 尚未產生
校正前後逐字稿
目前待解決問題與初步改善方向
- 完整校讀／語意翻譯仍待確認，保留可用逐字稿發布
ValueError('Canonical gate requires review; translation safely withheld')
- 自動摘要未完成
已先發布可跳轉的逐字稿版本；可稍後以現有逐字稿重新產生摘要。；note generator exit=1, timed_out=0
- Qwen3-ASR 未完成，需確認 qwen_asr_lab 環境與音訊後重試精進模式。