YT to Skill

About the project

A command-line tool that watches YouTube trading videos, starting with Chinese crypto channels, and turns each strategy into a skill: a set of instructions an AI coding assistant like Claude Code can install and follow, plus the key chart frames.

A lot of good trading knowledge is locked in long videos, often in Chinese. YT to Skill pulls it out. Give it a video, a playlist or a whole channel, and it skips the videos that have no real strategy, translates the rest, and writes each strategy as a skill an AI agent can install and follow.

It's built as a pipeline of 7 stages, each saved to disk, so a re-run picks up where it stopped and one bad video never breaks a batch. Version 0.1 works end to end; the last open step is checking the keyframe tuning on real trading videos.

How it works

1. 1

   Find the videos

   Paste a video, playlist or channel link.

2. 2

   Get the words

   Use the captions if they're good. If they're poor, transcribe with a Chinese speech model instead.

3. 3

   Skip the noise

   A free keyword check, then an AI check, drops videos with no actual strategy.

4. 4

   Translate

   AI translation with a 97-term Chinese-English trading glossary in the prompt, so slang comes out right.

5. 5

   Extract

   The strategy goes into a strict format: entries, exits, risk rules. Anything the video doesn't say is marked as needing a value.

6. 6

   Write the skill

   A SKILL.md with the steps, plus the key chart frames from the video.

Under the hood

Pipeline

Python stages for ingest, transcript, filter, translate, extract, skill and keyframes, each cached on disk.

Speech

yt-dlp for the videos, a caption-quality check, and faster-whisper for Chinese speech.

AI

Claude through OpenRouter for translation and extraction, a smaller model for filtering, with output checked against a schema.

Keyframes

Scene detection plus change checks for long chart scenes, with near-duplicates removed.

Tests

196 test functions across 14 files.

What I did

• Designed the 7-stage pipeline, with every stage cached so re-runs skip finished work
• Caption-quality checks with a Chinese speech-to-text fallback
• A two-step filter: keyword scoring first, then an AI check
• A 97-term Chinese-English trading glossary for translation
• Strategy extraction into a strict schema, marking any value the video doesn't give
• SKILL.md output that follows Claude Code's skill format, with assets and references
• Keyframe capture for long screen-recorded chart scenes
• Typed errors, so one failed video doesn't stop a batch, and 196 tests

What was new

• A trading glossary inside the translation prompt, for slang that general translation gets wrong
• Capturing a chart frame when the screen actually changes, not on a timer
• Output you can install as an agent skill, not just a summary

Problems I hit, and how I fixed them

1. 01

   Hashtag spam in descriptions made the keyword filter reject real strategy videos.

   Fix Title keywords now count three times as much.

2. 02

   The AI filter was told "when in doubt, not a strategy", so Chinese strategy videos were thrown out.

   Fix Added Chinese trading terms, and it leans towards strategy when price levels and long or short appear.

3. 03

   Long chart scenes produced a single keyframe.

   Fix Check every 3 seconds in scenes over 45 seconds, save a frame when the picture really changes, and remove near-duplicates more strictly.

4. 04

   The SKILL.md files didn't match Claude Code's skill format.

   Fix Rewrote the header and added an instruction preamble.

Screenshots

To be clear

The repo's commits are under my TermiX teammate Ashutosh Pradhan's GitHub account, where it lives. The work is mine. I built it with an AI coding agent working through a step-by-step plan.