youtube-to-ebook: The Anti-SaaS Pipeline for the Slow Web

How a local Python daemon uses Claude 3.5 Sonnet to transform chaotic video feeds into curated, magazine-grade EPUBs.

7 min read • View on GitHub • More from zarazhangrui

A chaotic web of film strips feeding into a mechanical printing press that outputs a perfectly bound book. This illustrates the transformation of noisy video media into quiet, structured text. — Converting the algorithmic feed into a finished, offline product.

Key Takeaways

youtube-to-ebook acts as a local-first editorial desk that converts video transcripts into polished EPUBs.
The system injects video descriptions into Claude 3.5 Sonnet to fix phonetic hallucinations and enforce a strict literary tone.
By running as a local macOS daemon using the Supadata API, it bypasses aggressive cloud IP blocks from YouTube.
The project prioritizes finished, readable offline products over raw Markdown knowledge graphs.

The Prompt is the Editor-in-Chief

Most transcription wrappers treat text as raw data. They summarize, extract, and dump words into a database. The youtube-to-ebook project takes a different approach. It acts as an automated editorial desk. Its goal is to produce a finished, literary product that you can read offline.

Convert YouTube videos into stylish EPUB ebooks, complete with polished transcripts and optional email delivery for effortless reading.

HoangTheElegant/youtube-to-ebook README, Project Documentation · HoangTheElegant/youtube-to-ebook

The most fascinating intellectual choice in the codebase is how it handles context. AI transcriptions are notorious for phonetic hallucinations. If a speaker mentions the math channel 3Blue1Brown, a raw transcript might read 'Mark Brown'. To solve this, the script passes the video description into Claude alongside the transcript to act as an anchor.

Two hands working on a manuscript. A mechanical hand writes with a pen while a human hand holds up a reference index card. This visualizes the use of a video description as ground-truth context to guide the LLM. — Context injection forces the LLM into the role of a fact-checking editor.

Surviving the Extraction War

Scraping YouTube is a brittle endeavor. Standard libraries are frequently rate-limited or blocked by Google. The project abandons native scraping in favor of the Supadata API. This offloads the proxy rotation and bot-detection bypass to a dedicated service.

It also employs a clever HTTP redirect hack to filter out YouTube Shorts. Instead of relying on flawed metadata, a simple requests.head call checks if the URL redirects. If it does, the script instantly discards the short-form content.

The isolated pipeline ensures atomicity. If a step fails, the local state prevents duplicate processing.

A Daemon, Not a Web App

This is deliberately not a cloud application. By running locally via macOS LaunchAgents, the tool avoids server costs and bypasses cloud IP blocks. It maintains a strict state using a local JSON file to ensure atomicity before emailing the final EPUB.

def mark_videos_processed(video_ids):
    processed = load_processed_videos()
    processed.update(video_ids)
    with open("processed_videos.json", "w") as f:
        json.dump(list(processed), f)

The Reader vs. The Note-Taker

The YouTube-to-text space is divided by philosophy. Tools like Ebrizzzz focus on Markdown for knowledge graphs and study notes. The youtube-to-ebook pipeline is built for the Slow Web reader who wants a finished, offline magazine delivered to their inbox.

Feature	youtube-to-ebook	Ebrizzzz
Target Output	EPUB & Email	Markdown
Primary AI	Claude 3.5 Sonnet	Gemini
Extraction Method	Supadata API	Native API
Execution Environment	Local Daemon	Desktop App
Core Philosophy	Slow Web Reading	Knowledge Base Ingestion