11:32Andy Diep · Tutorial
3 Tips for AI Editing in Claude Code + DaVinci Resolve
An unscripted walkthrough of the editing philosophy, and the actual Claude Code prompts, behind one creator's AI-assisted cutting workflow.
September 2ndThe argument in one line.
Claude Code can operate DaVinci Resolve directly through an MCP connector, but it edits from transcripts rather than intent, so pairing the raw transcript with the original script and running a dedicated audit pass is what actually catches leftover retakes and cuts manual cleanup time by roughly 80%.
Read if. Skip if.
READ IF YOU ARE…
- You already record raw solo video and manually scrub through takes to build a clean cut.
- You use Claude Code for coding work and are curious whether it can be pointed at non-coding software through MCP.
- You want a repeatable AI-assisted rough-cut workflow for DaVinci Resolve before doing a final manual pass.
SKIP IF…
- You edit in Premiere, Final Cut, or another NLE — this setup is specific to DaVinci Resolve Studio.
- You want a fully hands-off finished edit — the video is explicit that a human review pass is still required.
The full version, fast.
Claude Code can drive DaVinci Resolve Studio directly through the DaVinci Resolve MCP, letting a creator hand over a raw recording and get a cut back without touching the timeline. The setup needs DaVinci Resolve Studio with external scripting turned on and a paid Claude plan, connected with one install command. The key lesson: feeding the AI just the auto-generated transcript isn't enough, because speech-to-text engines clean up stumbles and restarts, hiding the very repetition an editor needs flagged. Supplying the original script alongside the transcript, then running a dedicated audit pass that cross-checks script, transcript, and timeline, caught 17 retakes a first pass missed and cut manual cleanup from two to three hours down to about thirty minutes.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:47
01 · I Let Claude Code Edit My Raw Video
Cold open debunking the idea that AI video editing needs a terminal, a programming language, or a dev background, and previewing that the process still needed a human audit.
00:47 – 01:25
02 · Connect Claude Code to DaVinci Resolve
Lists the requirements: DaVinci Resolve MCP, a paid Claude Code plan or API credits, the Claude Code desktop app, and DaVinci Resolve Studio for external scripting.
01:25 – 02:05
03 · Install the DaVinci Resolve MCP
Turns on external scripting in Resolve's preferences (System > General > Local), then installs the MCP with a single npm command pasted into Claude Code.
02:05 – 03:12
04 · Why Your Original Script Works Better
Loads the raw recording into the media pool and explains why a written script beats a raw transcript as Claude's reference, since the transcript contains every mistake, restart, and unused take.
03:12 – 04:08
05 · What Claude Got Wrong
Claude scores each take against the script, defaults to the later take on ties, and produces a longer but cleaner second cut; the first cut had over-trimmed footage and sliced real words.
04:08 – 05:07
06 · The 17-Retake Audit
A dedicated audit prompt cross-checks the script, transcript, and timeline, hunts repeated phrases within and across cuts, and surfaces 17 retakes still sitting in the edit.
05:07 – 05:55
07 · Why AI Transcripts Miss Editing Mistakes
Transcription tools smooth stumbles into readable text, hiding repetition; using script and transcript together gives Claude a better reference but still doesn't replace human review.
05:55 – 06:28
08 · From 3 Hours of Cleanup to 30 Minutes
Cleanup manually finished inside Resolve in about 30 minutes instead of the usual two to three hours, since the AI cut removed most of the repetitive work even though it wasn't a finished edit.
06:28 – 07:02
09 · Turn the Workflow Into a Claude Skill
Saves the one-page audit prompt as a reusable Claude skill and sets up a project holding the script, transcript, and timeline context for future videos.
Lines worth screenshotting.
- A raw recording's transcript captures every stumble and retake, but a written script only holds what the creator meant to say, giving an AI editor a cleaner target to cut toward.
- When two takes score identically against a script, defaulting to the later take works because a person's last attempt at a line is usually their most polished one.
- An AI edit that scores every take against a script found the correct take with 94% confidence in one case, but a tie at 88% still needed a fallback rule.
- A first AI cut over-trimmed footage and sliced real words at some cut points, running about a minute shorter than the version built with the original script as a guide.
- Speech-to-text engines are built to produce readable text, so they quietly erase the stumbles and restarts an editor needs to see in order to know a retake happened.
- A dedicated audit pass, cross-checking the transcript, the script, and the edited timeline, caught 17 retakes that two rounds of AI cutting had missed.
- Feeding an AI editor both the raw transcript and the original script gives it two references instead of one, catching mismatches neither document alone would reveal.
- A cleanup pass that normally took two to three hours dropped to about thirty minutes once an AI editor produced a strong first pass to work from.
- Saving a working prompt as a reusable skill turns a one-time experiment into a repeatable step that doesn't need to be re-explained on the next video.
- DaVinci Resolve requires the paid Studio version for external scripting, so the free version can't be controlled by an AI coding tool at all.
Give the AI both the script and the transcript.
WHAT TO LEARN
An AI video editor cuts from what a transcript literally captured, so pairing it with the original script it was written from is what actually catches retakes and mistakes a transcript-only pass misses.
01I Let Claude Code Edit My Raw Video
- Editing video with Claude Code doesn't require a terminal, a programming language, or a dev background — it works from plain-language prompts.
- The workflow ran three rounds of cuts, and even the final pass still needed a human audit to catch what the AI missed.
02Connect Claude Code to DaVinci Resolve
- This setup needs the DaVinci Resolve MCP, a paid Claude Code plan or API credits, and DaVinci Resolve Studio — the free version doesn't support external scripting.
- Claude Code doesn't require a coding terminal for this workflow; the desktop app is enough.
03Install the DaVinci Resolve MCP
- External scripting has to be manually turned on inside DaVinci Resolve's preferences (System > General) before any outside tool, AI or otherwise, can control it.
- The MCP installs with a single command pasted into Claude Code, which also handles missing dependencies automatically.
04Why Your Original Script Works Better
- A raw recording's transcript contains every mistake, restart, and unused take; a script only contains what the creator meant to say, giving the AI a cleaner target.
- Working from the full transcript alone, Claude sometimes kept lines that were never meant to be in the final cut.
05What Claude Got Wrong
- Claude scored every take against the script and picked the best match, defaulting to the later take on a tie — a rule that fit how the creator naturally records.
- The first cut over-trimmed the footage and sliced real words at some cut points; the script-guided second cut ran longer but was noticeably cleaner.
06The 17-Retake Audit
- A dedicated audit prompt cross-checked three sources: the original script, the full transcript, and the timeline Claude had already built.
- The audit searched for repeated phrases within a single clip and across neighboring cuts, and confirmed every script passage appeared exactly once — turning up 17 retakes that had survived two rounds of cutting.
07Why AI Transcripts Miss Editing Mistakes
- Speech-to-text transcription is built to output readable text, not a literal record, so it quietly smooths over the stumbles and restarts an editor needs to see.
- Using the script and the transcript together gave Claude a better reference than either alone, but it still didn't replace a final human review.
08From 3 Hours of Cleanup to 30 Minutes
- Manual cleanup that normally took two to three hours dropped to about thirty minutes once the AI-built second cut gave a strong starting point to work from.
09Turn the Workflow Into a Claude Skill
- The one-page audit prompt was saved as a reusable Claude skill, so it doesn't have to be rewritten or re-explained for the next video.
Terms worth knowing.
- DaVinci Resolve MCP
- A connector (built on the Model Context Protocol) that lets an AI coding tool like Claude Code send commands directly to DaVinci Resolve's editing timeline.
- External scripting
- A DaVinci Resolve Studio setting, off by default, that must be switched to Local before any outside program, including an AI tool, can control the app.
- MCP (Model Context Protocol)
- A standard that lets an AI assistant connect to and operate external software tools, in this case a professional video editor.
- Whisper
- An open-source speech-to-text model that Claude installed on its own to transcribe the raw recording's audio for the edit.
- Take-matching
- The step where the AI compares every recorded take of a line against the intended script to pick whichever take is the closest match.
Things they pointed at.
Lines you could clip.
00:35
“Claude's own audit found 17 retakes and mistakes still hiding inside the timeline. They survived because of a critical flaw in how AI handles audio transcripts.”
02:04
“A transcript contains everything you actually recorded. That includes mistakes, restarts, and unused takes. Your script contains what you intended to say.”
03:20
“My final attempt is usually my strongest attempt.”
05:07
“Transcription tools try to create readable text. They do not preserve every stumble exactly.”
05:43
“That cleanup took me about thirty minutes. Normally, this part takes me two to three hours.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
I just had Cloud Code edit a twenty two minute raw recording inside DaVinci Resolve. It removed hours of repetitive manual work without me writing a single line of code. Most people think Cloud Code is only a tool for software developers, but you can do so much more by simply talking to it and using commands from existing tutorials.
That's exactly what I did here. I copied one installation command and turned ClawCode into a video editor that cleaned up my raw footage inside DaVinci Resolve. It did what I wanted it to do, but it was not perfect.
After the second cut, Claude's own audit found 17 retakes and mistakes still hiding inside the timeline. They survived because of a critical flaw in how AI handles audio transcripts. In this video, I'll show you the exact setup I used to connect Claude code to Resolve, why it missed those retakes, and the one page prompt I now use to reduce my own cleanup from two to three hours to about thirty minutes.
It includes the instructions I developed through all the trial and error, so you don't have to repeat the same process. The setup itself is fairly simple. You need DaVinci Resolve MCP.
I link the GitHub page in the description. You also need access to Cloud Code. That requires a paid Cloud Plan or API credits.
You do not need to work inside a traditional coding terminal. I use Cloud Code through the desktop app. And obviously, you need DaVinci Resolve.
The studio version is required for external scripting. That scripting connection lets Cloud interact with Resolve. So let's start by enabling it.
Inside Resolve, open preferences. Choose system, then choose general. Find the external scripting using option.
Set it to local. After that, install the DaVinci Resolve NCP. The installation command is on screen now.
You can also find it in the GitHub installation walkthrough that I linked earlier. You only need the installation command. Paste that command into ClotCode, and it will handle the installation process.
If you're on Mac, that includes missing dependencies such as Node and Homebrew. Once the MCP was installed, Clot needed my raw recording. I placed the footage inside ClotResults media pool.
Klot also needed a reference for the intended final audio. If you have the original script, use that. Otherwise, you can use Resolve's transcript instead.
A transcript contains everything you actually recorded. That includes mistakes, restarts, and unused takes. Your script contains what you intended to say.
That gives Claude a clear baseline for editing. I discovered this during the first cut. Claude initially worked from the complete recording transcript.
That gave it all twenty two minutes of spoken material, but Claude could not always identify my intended version. It sometimes 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.
That made its decisions more precise. Claude also installed Whisper, downloaded a small speech model from Hugging Face, and installed the tool it needed to extract the audio, then analyzed the recording's audio and made the edits.
I did not do any research or install the tools myself. I provided the script and reviewed the results. When I noticed a problem, I described it to Claude.
Claude suggested an approach for fixing it. I approved the approach and let it continue. That is how the workflow progressed through a couple of iterations.
For the second cut, Claude matched my script against the raw recording, compared every passage with the available takes, then he selected the take matching the script best. When two takes matched equally, Claude chose the last one.
That actually made sense for how I normally record. My final attempt is usually my strongest attempt. The second timeline ran about ten minutes and thirty five seconds.
That was longer than the first cut, which ran about nine and a half minutes. The first cut had over trimmed the footage and sliced real words at several guest cut points. Points.
The opening was fixed. The second cut was much closer to what I wanted, but some problems were still there.
Claude missed several quick restarts. Sometimes I said a health word, then restarted. Some full phrases also appeared more than once.
Real quick before we continue, if this has been helpful so far, do me a favor, subscribe and like the video. It really helps grow the channel, and I genuinely appreciate your support. Alright.
Back to it. With those problems still in the timeline, I asked Clot to perform an extensive audit. I told it to compare three things, my original transcript, the transcript, and the timeline it created already.
Clot search for repeated phrases inside each clip. It also checked for repetition across neighboring cuts. When a phrase appeared twice, it kept the last version.
Then it compared the timeline against my script. Every script passage needed to appear exactly once. Anything repeated or missed needed to be flagged.
That audit found 17 remaining retakes. I saved that instructions as a single prompt. It fits in one page.
It tells Cloud which transcript to build, how to choose between takes, and exactly what to verify before calling a cut finished. The link is in the description.
These technical details came directly from Cloud after I asked what it had done and how. Again, my part in the workflow was minimal. I provided the original script, I described the problems I could hear, and I approved Cloud's proposed fixes.
That was it. So why did Claude let those retakes survive? It's because Claude relied heavily on transcription to understand the footage, but .
They do not preserve every stumble exactly. When I restarted a word quickly, the transcript often cleaned it up, sometimes displayed the intended phrase only once. That made the audio appear cleaner than it really was.
Repetition was missing from the transcript, Claude had less evidence that a cut was needed. That's why the original script improved the workflow.
The script showed what I intended to say. The transcript showed what the engine believed it heard. Using both gave Cloud a better reference, but it still didn't replace a final human review.
By the time Cloud completed the audit, I had already started cleaning the second cut manually, so I did not request a third timeline. I fixed the remaining issues inside the second cut in DaVinci Resolve. That cleanup took me about thirty minutes.
Normally, this part takes me two to three hours, so the result was still a major improvement for my workflow. Cloth did not produce a completely finished edit, but it gave me a strong starting point and removed hours of repetitive cleanup.
That was the real value of this experiment. It gave me a new workflow that I plan to continue using.
I saved the entire process as a Cloth skill. I also created a project for the workflow. The skill contains the instructions we developed.
That is the same one page prompt saved so I never have to paste it again. The project contains the context from this experiment. For my next video, I will not need to explain every single step.
I can begin with one prompt. Cloud can apply the rules from this project. Then I can review and finish the result in a matter of minutes.
One quick note about the Cloud DaVinci Resolve MCP. It also supports several additional capabilities. You can review those in its GitHub page I linked in the description.
Finally, I hope you found this useful. If so, help me grow the channel by liking the video and subscribing. I also just launched a free school community called AI Made Easy.
Inside, you will find the prompts and cloth skills I use in my videos. It's completely free, and the link is in the description. Thanks for watching, and I'll see you in the next
The Hook
The bait, then the rug-pull.
A creator handed Claude Code twenty-two minutes of raw footage and let it run the DaVinci Resolve timeline start to finish. Two automated cuts later, a dedicated audit still turned up seventeen retakes hiding in the edit, and the reason traces back to a blind spot in how AI transcribes speech.
Named ideas worth stealing.
04:08concept
Script + Transcript + Timeline Audit
- Original script
- Raw transcript
- Existing timeline
A verification prompt that cross-checks three sources against each other: it hunts repeated phrases inside a clip and across neighboring cuts (keeping the last version when a phrase repeats), then confirms every script passage appears in the timeline exactly once, flagging anything missing or duplicated.
Steal forany AI-assisted rough cut where the first pass needs a second, independent verification step before calling it done
03:17concept
Take-Matching Tiebreak Rule
- Score every take against the script
- Pick the highest match
- On a tie, pick the last take recorded
When two takes score identically against the intended script, the rule defaults to whichever take was recorded last, on the logic that a person's final attempt at a line is usually their strongest.
Steal forany automated take-selection step in a solo-recorded video workflow
How they asked for the click.
VERBAL ASK
06:43subscribe
“Like and subscribe, it grows the channel. I also just launched a free community called AI Made Easy — the prompts and Claude skills are inside.”
Delivered direct to camera at the very end with on-screen LIKE/SUBSCRIBE buttons and a community plug; a softer version of the same ask also appears mid-video around 3:54.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
