Editor.
I made Claude Code edit a 22-minute video inside DaVinci Resolve:
PROMPT: https://www.skool.com/ai-made-easy-5911/classroom
I forced Claude Code to video edit a 22-minute raw recording inside DaVinci Resolve.
It cut out hours of manual work. I never wrote a single line of code.
The full walkthrough is in the video above. Here’s the written version.
What you need.
Most people think Claude Code is a tool for software developers.
But you can connect it through MCP, Anthropic’s system for hooking AI up to other apps. Do that, and Claude controls your timeline for you.
Three things:
The DaVinci Resolve MCP. Free and open source: github.com/samuelgursky/davinci-resolve-mcp
Claude Code. Needs a paid Claude plan or API credits.
DaVinci Resolve Studio. External scripting is Studio only.
Turn on scripting.
That scripting connection is what lets Claude interact with Resolve, so start there.
Inside Resolve, open Preferences. Choose System. Then General.
Find External scripting using and set it to Local.
Install the MCP.
You only need the command:
npm install -g davinci-resolve-mcp
Paste it directly into Claude Code and it handles the installation for you. That includes missing dependencies like Node and Homebrew.
Give Claude the footage and the script.
Once the MCP was installed, Claude needed my raw recording. I put the footage in Resolve’s Media Pool.
It also needed a reference for the intended final video.
That reference can be a transcript. Setup guides recommend using Resolve’s transcription feature, and I did that on my first attempt.
But your own video script works much better.
A transcript contains everything you actually recorded, including mistakes, restarts, and unused takes. Your script contains what you intended to say.
I found this out on the first cut. Claude worked from the full recording transcript, so it had all twenty-two minutes of spoken material. It couldn’t always identify my intended version, and it kept lines I never meant to use.
When I supplied the original script, the cuts improved. Claude could compare the recording against my intended narration, and its decisions got more precise.
If you already have a written script, use it. If not, Resolve’s transcript will do.
What Claude did on its own.
It installed Whisper, the transcription tool. It installed the tool it needed to extract the audio. And it downloaded a small speech model from Hugging Face.
Then it analyzed the footage’s actual audio and made its edits.
I did not write any code. I did not research or install any of those tools myself.
I provided the script and reviewed the results. When I noticed a problem, I described it. Claude suggested an approach. I approved it and let it continue.
That’s how the workflow moved through several iterations.
How it picked the takes.
For the final cut, Claude matched my script against the raw recording. It compared every passage with the available takes, then selected the one matching the script best.
When two takes matched equally, it chose the last one. That works for how I record. My final attempt is usually my strongest.
The second timeline came in at ten thirty-five. The first came back at nine and a half.
Shorter was not better. That first cut was over-cutting, slicing real words in half at guessed cut points. The second one fixed the opening and landed much closer to what I wanted.
Then I found 17 more retakes.
Some problems were still in there. Claude missed several quick restarts. Sometimes I said half a word, then started over. Some full phrases appeared more than once.
So I asked Claude to run an extensive audit, comparing three things: my original script, the transcript, and the timeline it had already created.
It searched for repeated phrases inside each clip. It checked repetition across neighbouring cuts. When a phrase appeared twice, it kept the last version.
Then it compared the timeline against my script. Every passage needed to appear exactly once. Anything repeated or missing got flagged.
That audit found seventeen remaining retakes.
I saved that instruction as a single prompt. It fits on one page. It tells Claude which transcript to build, how to choose between takes, and exactly what to verify before calling a cut finished. It’s linked under the video.
Why those retakes survived.
This was the important question, so I asked Claude what it had done and how. All of these technical details came from it directly.
Claude relied heavily on transcription to understand the footage. But transcription tools try to create readable text. They do not preserve every stumble exactly.
When I restarted a word quickly, the transcript often cleaned it up and showed the intended phrase only once. That made the audio look cleaner than it really was.
The repetition was missing from the transcript. So Claude had less evidence that a cut was needed.
That’s exactly why the original script improved things. The script showed what I intended to say. The transcript showed what the engine believed it heard. Using both gave Claude a better reference.
It still didn’t replace a final human review.
What it cost me.
By the time Claude finished the audit, I’d already started cleaning the second cut by hand, so I never asked for a third timeline. I fixed the rest inside the second cut.
That cleanup took me about thirty minutes.
Normally this part takes me two or three hours.
Claude did not produce a completely finished edit. But it gave me a strong starting point and took hours of manual work off my plate.
That’s the real value I found here. An experiment that turned into a workflow I’ll keep using.
So I saved it.
I saved the entire process as a Claude skill. I also created a project for the workflow.
The skill holds the instructions we developed, the same one-page prompt, saved so I never paste it again. The project holds the context from this experiment.
For my next video I won’t explain any of this again. I’ll start with one prompt, Claude will apply the rules from the project, and I’ll review and finish in minutes.
One last note on the MCP. It supports several other capabilities beyond editing. They’re on the GitHub page.