Edit Shorts With Claude Code: Word Stickers, Text Behind You (Free Kit)

Install Claude Code from the official docs (or use the desktop app’s Code tab), sign in once, and check it runs:

claude --version
node --version
ffmpeg -version
pip install faster-whisper

Don’t want to install tools by hand? Skip ahead: the first request in Step 2 tells Claude to check and install what’s missing. It will ask before each command; inside the kit folder it’s safe to allow.

Download claude-auto-edit-kit.zip, unzip it, and put two files in the folder:

• talk.mp4: your vertical talking clip with its own sound
• script.txt: exactly what you said, word for word

Filming tips that save fixes: shoot vertical; leave space above your head and below your chest (captions need room); speak clearly; keep the clip’s own sound (a dubbed voice makes the lips and captions drift). Aim for about 75 words in 30 s.

Open Claude Code in the kit folder (cd into it, then claude; or pick the folder in the desktop app) and paste:

This folder is the claude-auto-edit-kit. Read README.md first and follow it.
talk.mp4 is my vertical talking clip and script.txt is exactly what I said.

1. Check Node, ffmpeg, Google Chrome and Python with faster-whisper. Install whatever is missing (ask before anything global).
2. Run python prep.py, then compare words.json with script.txt and fix any misheard words.
3. In build.py, change every at("...") anchor to the matching word in MY script, keeping the scene order.
   Replace the on-screen texts (the big word behind me, the typed line, the chat lines, the end card) with my content.
4. Run python build.py and python mix.py, then npx hyperframes check, and fix anything it reports.
5. Grab one still frame per scene and make sure no caption or sticker covers my face. Fix and re-check.
6. Render a quick draft: npx hyperframes render -q draft -o renders/draft.mp4. Show me the file. High quality only after I approve.

Claude installs what’s missing, transcribes, retimes the scenes to your words, checks still frames, and renders a draft. Rendering draws every frame in a browser, so it takes a few minutes. That’s why the first pass is a quick draft.

Check three things in the draft: captions appear as you say each word, nothing covers your face, and none of our sample text (Claude / Adobe) is left on screen.

Same conversation. Point at the moment by time or by line, one or two fixes at a time, and add partial fix only so Claude doesn’t rebuild everything. A screenshot of the bad frame helps. Examples:

At 3.2 s the caption covers my mouth. Move that caption block up 200px, partial fix only.

The stickers pop too fast. Make each pop 0.3 s and stagger them by 0.12 s.

Change the big word behind me from CLAUDE to my channel name, 6 letters max.

The green scene feels empty. Make the window and the step chips 20% bigger.

Where our version needed the most work: captions covering the face in the first draft, a misread word in the voice track, and too many effects stacked at once. Check those first.

Lock this version. Render it with npx hyperframes render -q delivery -o renders/final.mp4.
Keep my original voice. If assets/bgm.mp3 exists, mix it 15 dB under my voice and fade out the last second.

Only use music you have the rights to. If you want a trending song, leave it out here and add it inside Instagram or TikTok when you post.