How to Edit a Screen Demo with Claude Code on Mac

You already pay for Claude Code, and your demo recordings still need an hour of cutting. Claude Code can do that first pass if the editor gives it real editing tools instead of a screenshot to click.

The short answer: record the demo in ScreenKite, start Claude Code in the folder that holds the project, and ask for a transcript cut with the plan first. Claude Code returns the list of cuts, you read it, and the cuts land when the agent applies them.

What you need

• A Mac on macOS 15 or later, Apple Silicon or Intel.
• ScreenKite, signed in with your account.
• Claude Code installed from Anthropic's documentation.
• A recording. Any screen demo with a voice track works.

How to edit a demo with Claude Code in the terminal

Step 1. Record or open the demo. Record in ScreenKite, or open an existing project. ScreenKite saves each project as a bundle on your disk.

Step 2. Set up transcription once. The agent cuts from the words, so it needs a transcript.

1. Open Settings, Transcription, then the Word-Level tab.
2. On Apple Silicon, set the provider to Local, pick a WhisperKit model and click Download Selected Model. The audio never leaves your Mac.
3. On Intel, set ElevenLabs cloud transcription on the same screen.

Step 3. Start Claude Code next to the project.

cd ~/Desktop
claude

Step 4. Ask for the plan, not the cut.

Open ~/Desktop/Recording.skbundle and do a transcript cut.
Remove false starts, filler words and dead air.
Show me the cut list first. Do not apply anything until I say yes.

Claude Code returns each cut with its time range and a reason, such as a false start or a filler word, plus the new running time. The timeline has not changed yet.

Step 5. Read the list, then apply. Check the first and last cut, where a clipped word hurts most. Reply "yes", and Claude Code sends the second call that applies the cuts. Ask for changes first if a cut looks wrong.

Step 6. Check and export. Scrub the result in ScreenKite, then export a local MP4.

What this does well:

• Removes false starts, fillers and dead air from a talk track in one pass.
• Keeps every cut as an ordinary edit on the ScreenKite timeline, so you can undo it.
• Leaves the project on your Mac.

What to watch:

• Claude Code in a terminal is not gated. It can send the plan call and the apply call itself, so the "show me first" line in your prompt matters.

Method 2: the chat panel inside ScreenKite

ScreenKite has a built-in AI chat panel with the same editing tools, and no terminal.

1. Open the project in ScreenKite.
2. Open the AI chat panel in the editor.
3. Ask for the same transcript cut.

The chat panel gates each tool on its risk tier. Pick the approval mode that stops before a cut, and you approve each plan in the panel.

Go further with skills

ScreenKite publishes prompt bundles for Claude Code in its awesome-ai-video-editing repository. Install them once:

npx skills add ScreenKite/awesome-ai-video-editing

The B-roll skill maps your cut transcript to beats, proposes a set of visual slots, renders them and places each one as a scene layout on the timeline. Call it by name in Claude Code.

What ScreenKite gives Claude Code

• A real multi-clip timeline. The agent cuts, trims and lays out scenes on the same timeline you use.
• A plan for the edits that change meaning. Cuts, filler and dead-air removal, chapters, caption imports and voiceover swaps show their plan before they apply.
• Your own agent plan. Claude Code bills to the plan you already have. ScreenKite adds no AI fee.
• One app from recording to export. No round trip through a second editor.

One product demo once took me an afternoon. Screen Studio was slow and thin on editing, so the clip went to Final Cut. I fixed the cuts, the animations broke, and I exported again. Two tools, two payments, one afternoon gone. With Claude Code on the ScreenKite timeline, the first pass is a prompt.

How it is built

• 83 purpose-built editing tools per window. Each open editor window serves them over local loopback, one server per window, shaped by tens of thousands of hours of video-editing practice. Claude Code calls editing operations, not pixels on a screen.
• The same mutation coordinator as the buttons. A tool edit registers undo exactly like a click, so an agent pass is as reversible as your own.
• Propose, then commit, on named actions. A timeline cut, the silence, word and dead-air cuts, generated chapters, caption import and patch, and voiceover and camera replacement return a plan on the first call and mutate nothing. A second call applies it. Removing a track or an asset applies at once.
• A revision guard. Each agent edit carries the project revision it expects, which protects the timeline from an edit made against a stale copy.
• Transcription on the Mac. WhisperKit runs on Apple Silicon, so the talk track is transcribed without an upload.

Tips for a cleaner first pass

• Name what to keep. "Keep the pause before the price reveal" saves a round of fixes.
• Cut in passes. Ask for fillers first, then pacing. Short plans are easier to read.
• Proofread product names. Speech recognition mishears brand names. Ask Claude Code to fix them before it builds captions.
• Check the ends. Read the first and last cut in every plan.
• Save a prompt file. Keep your house rules in a text file and point Claude Code at it each session.

Facts

Common questions

Can Claude Code edit video on a Mac?

Yes, when the editor exposes editing tools. In ScreenKite, Claude Code calls 83 editing tools on the open project, so it cuts, trims and lays out scenes on a real timeline.

Does Claude Code wait for my approval before it cuts?

Not by itself. A cut returns its plan first and applies on a second call, and Claude Code in a terminal can send both. Tell it to show the plan and wait, or use the chat panel inside ScreenKite, which gates on each tool's risk tier.

Do I pay ScreenKite for the AI?

No. Claude Code bills to the plan you already pay for. The ScreenKite AI fee is $0.

Does my recording get uploaded?

The project stays on your Mac. On Apple Silicon, transcription runs locally with WhisperKit. On Intel, transcription uses ElevenLabs, which you turn on in Settings. Claude Code sends your prompts and the text it reads to Anthropic, as it does in any project.

Can I undo what Claude Code did?

Yes. Agent edits register undo like a click. A timeline cut from the command line cannot be undone from the command line, so read the plan before you apply it, or undo in the editor.

Which other agents work the same way?

ChatGPT Codex, Antigravity, Grok Build and Kimi use the same tools and the same plan-first calls.

Conclusion

Record the demo, start Claude Code next to it, and ask for the plan first. ScreenKite gives the agent you already pay for a real timeline, for one payment.

Try ScreenKite