Youtube Hindi To Roman Transcript Skill

YouTube Hindi → Roman Transcript (Claude Agent Skill)

Turn a Hindi or Hinglish YouTube video — or an entire course playlist — into two clean text files: an exact line-by-line **Devanagari** transcript, and a readable **Roman Hindi (Hinglish)** version.

**This repo ships a procedure, not a program.** It contains one file, `SKILL.md`. There is nothing to install and no CLI to run. The skill teaches a Claude agent how to build three small Python scripts *inside your own project*, then how to run and refine them. That is deliberate — see [How it works](#how-it-works).

The problem

You have found a long Hindi YouTube video — a lecture, a multi-hour course, a tutorial — and you want an AI model to actually *read* it, so you can question it, summarise it, or learn from it.

But there is no transcript to hand it.

Four things stand in the way:

1. Feeding the video itself is expensive. Processing hours of frames costs far more than reading the words that were spoken — enough that transcribing a whole course stays theoretical instead of something you just do.
2. The captions that do exist come back in Devanagari. Many fluent Hindi speakers read and type in Roman script; a Devanagari wall of text is slow to read, or unreadable, for exactly the people who understand the content best.
3. Naive transliteration produces gibberish. Convert मतलब character by character and you get matalaba, not matlab. And because speech-to-text writes spoken English words in Devanagari too, range comes back as renj and consider as kansidar — every few lines reads broken.
4. Devanagari is expensive to feed back to a model. Mainstream tokenizers are trained on overwhelmingly English text, so Indic scripts fragment into many small pieces and cost noticeably more tokens for the same meaning (Tokenizer Tax, 2026).

The useful fact underneath all this: **a long Hindi video almost always has