Unpack it into your Claude Code skills folder:
# Windows (Git Bash) mkdir -p ~/.claude/skills tar -xzf chalk.tar.gz -C ~/.claude/skills # macOS / Linux mkdir -p ~/.claude/skills && tar -xzf chalk.tar.gz -C ~/.claude/skills
Restart Claude Code. Then just say “run Chalk” or “make a training video of this workflow” and it takes over.
chalk should be listed.| Step | Command |
|---|---|
| 1. Scaffold | python new_deck.py <folder> "Video Title" |
| 2. Write the deck | Edit slides.html — one .slide div per slide |
| 3. Write narration | Edit narration.json — one string per slide, same order |
| 4. Build | python make.py "Output-Name" |
| 5. Publish | python deploy.py <slug> "Page Title" |
In practice Claude writes steps 2 and 3 for you from the conversation you have just had. Your job is to check the content is true.
Teach the workflow, not the click path. A good training video explains why a step exists and what breaks without it. The videos that work carry the real rules we learned the hard way — not a list of buttons.
The build stops if they differ. If they drifted and you forced past it, every slide after the mismatch gets the wrong audio — and it looks fine until you watch it.
Finished segments are cached by duration, so editing a slide image without changing its narration reuses the old one. Delete that segment and rebuild:
rm segments/seg-15.mp4 python make.py "Output-Name"
The full list is in reference/GOTCHAS.md inside the skill. Read it before debugging.