The argument in one line.
Claude Code can drive DaVinci Resolve through an MCP server well enough to ingest footage, transcribe it, and assemble a rough first-pass cut, but deciding which cuts actually matter still has to come from the editor.
Read if. Skip if.
READ IF YOU ARE…
- A solo creator paying freelance editors weekly for a channel that isn't yet profitable and wants to cut that recurring cost.
- Someone comfortable cloning a GitHub repo and running terminal setup steps inside Claude Code without hand-holding.
- An editor curious whether an AI agent can operate a real NLE like DaVinci Resolve, not just generate captions or short clips.
SKIP IF…
- You want a fully polished edit handed back with zero manual cleanup — dead air and mistimed cuts still need a human pass.
- You have no script or reference doc to feed the agent; this workflow depends on a structured source document to cut against.
The full version, fast.
Paying $1,300/month for freelance editors on a channel making $0 forced a rethink, so the creator connected Claude Code to DaVinci Resolve using an open-source MCP server plus the Superpowers plugin's brainstorm skill. Given a footage folder and a script reference doc, Claude opens Resolve, transcribes each clip with WhisperX, and assembles a first-pass timeline on its own. Tests show it can follow relative marker instructions (move this pink clip to that marker) but its dead-space detection is inconsistent — it misses obvious gaps and flags irrelevant ones. The conclusion: let the agent handle repetitive assembly while you keep doing the judgment calls yourself, treating it like a junior editor you haven't fully trusted yet.
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 – 00:38
01 · The cost problem
States the problem that drove the build: paying about $350/week ($1,300/month) to three freelance editors on a channel earning $0, forcing a choice between finding an alternative or stopping uploads.
00:38 – 01:22
02 · What agent-editing means here
Defines the workflow as an editing assistant that produces a first-pass assembly cut while the human keeps final say, aimed at solo creators buried in repetitive timeline work.
01:22 – 02:31
03 · The DaVinci Resolve MCP repo
Introduces the open-source DaVinci Resolve MCP server (by Samuel Gursky) that lets a coding agent like Claude Code or Codex drive DaVinci Resolve directly.
02:31 – 03:29
04 · Running the setup plan
Creates a local project folder, pastes the MCP repo link into Claude Code, and runs it in plan mode at high reasoning so the agent scans the folder, drafts an install plan (clone, isolated venv, wire-up, verification), then executes it.
03:29 – 05:08
05 · Installing the Superpowers plugin
Adds obra's Superpowers plugin — either by pasting a GitHub prompt or through Claude's plugin browser — to unlock extra skills; the brainstorm skill is the one actually used later.
05:08 – 06:45
06 · Starting a session with a structured prompt
Opens the MCP folder as a fresh Claude Code session and hands it a template prompt: footage path, a script/reference doc (a Notion page), a scoped instruction ('just the first two minutes'), and an explicit call to the brainstorm skill in plan mode.
06:45 – 08:22
07 · Choosing Opus vs. Sonnet
Runs the brainstorm/plan step on Opus at high reasoning for quality, then switches to Sonnet for the execution pass to save tokens, noting high-reasoning Opus burns noticeably more.
08:22 – 09:07
08 · Watching the agent work autonomously
The agent opens Resolve, creates a new 29.97fps project timeline, and ingests three clips on its own — assembling beats 0 through 2 into a fully pre-cut sequence with no further prompting.
09:07 – 10:45
09 · The marker relocation test
Manually recolors a clip pink, places a cream-colored marker, and asks Claude to move the pink clip to the marker's position and delete the red markers — testing relative, color-coded references. It succeeds, but detaches adjacent clips as a side effect.
10:45 – 12:58
10 · How the pipeline actually works
Explains the mechanism: since AI can't 'watch' video like a human, the agent transcribes each clip with WhisperX, analyzes word-level timing, and builds per-clip alignment files it uses to assemble cuts — plus a warning to hand-hold it fully until you trust it, like onboarding a junior editor.
12:58 – 15:35
11 · The dead-space marking test
Asks the agent to mark every dead-space gap in the timeline with a red marker for review instead of cutting automatically, then finds the marks inconsistent — missing obvious gaps while flagging spots that don't matter.
15:35 – 16:43
12 · Doing the fine polish by hand
For small, fast fixes — trimming a lingering gap, nudging a clip to land on motion — doing it manually in Resolve is faster than prompting the agent; the two run in parallel while the agent works elsewhere.
16:43 – 17:37
13 · AI does the repetition, you keep the judgment
Closes on the thesis: the agent is good at specialist, repetitive work — cuts, assembly — but can't yet decide which pieces matter; that judgment call is the skill worth protecting and building.
Lines worth screenshotting.
- The creator was spending $350/week — about $1,300/month — on three freelance editors for a channel making $0 in revenue.
- The setup treats Claude as an editing assistant that handles first-pass assembly while the human retains final say over every cut.
- The DaVinci Resolve MCP server (built by Samuel Gursky) is what lets a coding agent like Claude Code or Codex actually operate DaVinci Resolve.
- AI agents can't watch video the way humans do — they transcribe audio into words and analyze frames separately, then reconstruct edits from that text layer.
- WhisperX is the tool Claude uses under the hood to transcribe footage and build word-level alignment for each clip.
- Running the initial brainstorm/plan step on Opus at high reasoning costs more tokens, so the creator switches to Sonnet for the execution pass to save budget.
- In a live test, the agent correctly moved a manually-recolored pink clip to a cream-colored marker's position, but it also detached the clip from its neighbors as a side effect.
- Asked to mark every dead-space gap in the timeline for review, the agent's markers were inconsistent — missing gaps the creator actually cared about and flagging spots that didn't need attention.
- The creator compares trusting an AI editor to trusting a brand-new junior editor: you don't hand over unsupervised first cuts until you've watched it work and built confidence.
- For small, obvious fixes — trimming a short gap, nudging a clip to land on motion — doing it manually in Resolve is faster than writing a prompt for it.
- The workflow runs in parallel, not sequentially: while the agent chews through repetitive assembly, the creator works on packaging, animations, and sound elsewhere.
- The core thesis: AI can only replace you if you're doing the exact same repetitive task it does — the judgment about which pieces matter is the skill worth protecting.
Let the agent do the repetitive assembly, keep the judgment calls yourself.
WHAT TO LEARN
An AI agent wired into a real NLE can genuinely do first-pass assembly work, but its marker logic and dead-space detection are unreliable enough that you still have to review every cut it makes.
01The cost problem
- Weigh the recurring cost of freelance editing against the setup cost of an agent-editing pipeline — $1,300/month on editors is a real number to compare against.
02What agent-editing means here
- An agent editor is only useful if you keep final review on every cut; treat its output as a first pass, not a delivery-ready edit.
03The DaVinci Resolve MCP repo
- MCP servers are what let a coding agent operate real creative software (here, DaVinci Resolve) rather than just generate text or clips.
04Running the setup plan
- Plan mode matters: seeing the agent's proposed steps before it executes catches problems before they're baked into your project file.
06Starting a session with a structured prompt
- Give the agent a scoped instruction (a time range, a chapter) instead of the whole video — smaller scope means fewer compounding errors to review.
07Choosing Opus vs. Sonnet
- Running the planning/brainstorm pass on a stronger, more expensive model and the execution pass on a cheaper one is a reasonable way to control token cost.
09The marker relocation test
- AI agents don't 'watch' video — they transcribe audio to text and analyze frames separately, so anything that depends on visual-only meaning may not translate cleanly.
- Test an agent's spatial/relative reasoning (move this to that marker) before trusting it with judgment calls like which silences to cut.
11The dead-space marking test
- Automatic dead-space detection is not yet reliable enough to run unsupervised — expect to manually re-check and often manually fix pacing gaps yourself.
12Doing the fine polish by hand
- The efficiency gain isn't 'don't do anything' — it's working in parallel: let the agent grind through repetitive cuts while you do packaging, sound, or the next piece of content.
13AI does the repetition, you keep the judgment
- The transferable principle beyond editing: AI replaces you only on tasks that are pure repetition; the skill worth building is the judgment about what actually matters.
Terms worth knowing.
- MCP (Model Context Protocol)
- A protocol that lets an AI agent like Claude Code call into external software — here, a DaVinci Resolve MCP server exposes editing actions (import, cut, marker, timeline moves) the agent can invoke directly.
- DaVinci Resolve
- A professional video editing application with a full timeline, color, and audio toolset, used here as the NLE the AI agent is being taught to operate.
- WhisperX
- A speech-to-text tool that transcribes audio into word-level timed text, which the agent uses to understand and align footage it otherwise can't 'watch' like a human.
- Superpowers plugin
- A community Claude Code plugin (by obra) that adds extra skills to the agent, including a 'brainstorm' skill used here to plan the edit before executing it.
- Beat
- In this workflow, a labeled chunk of the assembled timeline (e.g. beat 0, beat 1) representing one pre-cut segment of footage the agent has already assembled.
- Plan mode
- A Claude Code setting where the agent drafts and shows a proposed sequence of actions before executing anything, letting the user review the approach first.
Things they pointed at.
11:14toolWhisperX
06:15toolNotion
Lines you could clip.
00:09
“I only built this because I was spending about $350 a week on three freelance editors, roughly $1,300 a month, on a channel that was making $0.”
12:38
“Would you trust a junior editor or a brand new editor to your team with the first cut without giving them any directions at all?”
16:18
“AI empowers you. It can only replace you if you're doing the exact same thing it does.”
17:17
“Right now, it does the specialist work. The cuts, the assembly, the repetitive stuff. What it can't do yet is manage that whole thing. It doesn't know which pieces actually matter.”
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.
analogy
00:00So watch as populates a whole entire timeline by itself. That's Claude editing real footage in DaVinci Resolve by itself.
00:09And I only built this because I was spending about $350 a week on three freelance editors, roughly $1,300 a month on a channel that was making $0.
00:19It was a brand new channel with no money coming in. I remember doing that math because I'm Asian and just being like, okay, I either figure something else out or I stop uploading. So I set up Claude as my editor and I'm gonna show you the exact step in this video from start to finish because when I first sat down to try this, I genuinely thought it was going to be a waste of time, but now I prefer this over the traditional method of editing.
00:44So what I built here is basically an editing assistant. You set it up, it handles your first pass, assembly, cuts, timeline, and you come in and you finish it. And the good thing is that you still have the final call.
00:55Look, I had my start in editing eight years now. I don't mind it. I just got tired of the monotonous cuts that eat your whole afternoon.
01:02And especially if you're a solo creator buried in your timeline, this speeds that up dramatically. If you're newer and you actually want to watch the process happen in Resolve, you can learn from it as well.
01:13You don't have to hand everything to AI. This just takes that repetitive part off your plate. So right here, we have the DaVinci Resolve MCP by Samuel Gursky.
01:22This has all the skills that you would need in order to drive your agent, Cloud Code or Codex or whoever you're using in order to take control of your DaVinci Resolve editor. This has a lot of information and if you've never seen something like this before, this is a GitHub page.
01:39It's basically code base so that the editors or your agent can understand how to manipulate DaVinci Resolve the way that you would, except this has all been codified. You'll take this link here, you will copy, and then we'll move into your agent. With that said, what you want to do is you wanna go down here, open up a new folder, and scroll down to your directory.
02:01So your directory will be in here. It has a little house icon right next to it. Name it DaVinci Resolve MCP.
02:11I'm just gonna do this. We'll open it and you should see this right here. Now this is a local folder on your computer.
02:18Now I'm gonna tell it something like, help me set up this repo and we will drop in this link here.
02:26Do plan, keep on high, and press enter.
02:31And I'll let this run by itself. It will help me set up the entire repo. I'll show you the folder right at the very end.
02:36So it will go through and search your folder to see if it has any files or duplicates. What it does have with that, it creates us for us a plan. So it checks all of our files, make sure that we have the correct files and this is the step that it takes.
02:51So it's gonna clone it, it creates an isolated virtual, install the files in there, it wires it up, tell the user to check the resolve setting, make sure that your files are correct, and then verify it. I'm gonna go ahead and run this now, and by the end, pretty much you'll be following me step for step.
03:08Go ahead and put it on accept and then auto mode as well. Everything finished and over here, this is our folder.
03:17You can see the path down here. Oh, my goodness. 10 gigs left.
03:22Gonna have to clear out some things very soon. All the files that you saw from that GitHub repo is now within our computer. So from here, this is where we're gonna work on from now on whenever you start your Cloud Code session.
03:34But before we do any of those things, I want you to install the superpower skills. We'll copy and paste this github./obra/superpowers, and if you don't know what superpower does for you, it makes your agent way smarter than it really should be.
03:52It's a very great plugin. I use it for almost everything. I always use the brainstorm skill, but there's a whole host of other skills as well that I haven't really explored.
04:01It's just the brainstorm skill is so good in what it does. You'll paste in the prompt. I'm going to erase this part, but you get you'll keep this part in.
04:12Okay? Actually, you know what? You're gonna have to install the superpower skills before.
04:16So just paste this prompt into a new chat, run it, and once that's done, then you can run your original prompt. So I'm gonna go ahead and run this, or you can install superpowers from here as well. This is you go into your directory, go into your plugins, go to browse, type in superpowers and go to your partner.
04:37It's also right here. So you can install from right here. This is connected from Anthropic by the way, so very trustworthy, or you can do it from here.
04:46I would be careful about installing random things on GitHub. Sometimes, some of these things might be compromised, whereas if you install it from the Claude plugins itself, a little bit more well, I would say a lot more trustworthy.
05:00The next step what you're gonna do is you're gonna create a new session. Okay? Once your chat is done setting up all the folders, you would now have your folder ready in order to edit.
05:11I am going to click here. You can do command n or click on new session here and that will put you into a new session. You go down here, click on where this folder is at and do open folder, browse to where it says DaVinci Resolve MCP or whatever you name your photo earlier and press open from there.
05:29Now you're in your DaVinci Resolve MCP and we'll type a quick prompt. Hey, I'm starting on a new project called AI edits my videos for me.
05:39I am going to put in the link to the footage and the link to the script, the Notion page for you to reference off of to edit and cut.
05:50Let's do the first two minutes of the video. Use the brainstorm skill. We'll put this on to plan.
05:59Now I am going to pull my footage right here. We're gonna copy the file path or you can drag in the folder as well.
06:08Either way works. I do it this way though, and I'll go ahead and put in my Notion.
06:15Honestly, this is where I just keep all of my scripts. I keep everything inside of Notion, but if you have a centralized place where you can set up, it could be your notes, Google Docs, whatever the case is.
06:26Right? Wherever you keep your script organized at, paste it in here. I am going to use Sonnet because actually, we're gonna use Opus in order to plan and brainstorm out the video first.
06:40Once he gets all that done, we'll switch over to Sonnet five in order to save the token. This isn't the best of both worlds. You can use Fable two, but like I mentioned earlier, it comes with a lot of token consumption.
06:53So be aware of that fact. So do Opus. We'll keep it on high.
06:57If you're running a little bit low on tokens or you have $20 plan, you can also use medium as well. High is just more of a luxurious thing to do.
07:06Yeah. Here's my DaVinci Resolve. It's currently running through its code.
07:11It's gonna start a new file soon. Let's see what it's currently doing. So right here, you can see it ran the command to open DaVinci Resolve, and it's doing all of these at once.
07:21Okay. Now, the agent it did this by itself, by the way. It opened up Resolve, created a file called a I d r demo first two minutes.
07:29It ingested our footage right here. So we have three three clips. There's more footage to this video, but as a demonstration, we're only going to do the first two minutes.
07:39I don't wanna edit the whole entire video in this session. So you we can see what it's doing on the back end right now. It loaded up the environment.
07:48It imported C22888990, it created a project timeline at 2997, and now it's doing its other agent things.
07:59And so this part, you're going to see the agent pretty much just populate everything all at once. So this is the first clip. In a moment here, it'll start loading the second clip.
08:11There you go. We're seeing the next beat now. This is beat one, so that first one that's first session beat zero, beat one, and beat two has everything precut already.
08:22So we have all of our footage here. The next thing that I want to do, which I'm pretty curious about, I haven't tried this before yet.
08:30I'm gonna take this clip here. We're gonna change it to pink and we're gonna see if the AI recognizes or Claude recognize that this clip is pink. We wanna move this pink clip all the way to this marker.
08:48Let's change this marker to cream.
08:54Move the pink clip to here. We just want to see whether it can do things like this or not because imagine you have you're trying to describe something.
09:07You're trying to say, hey, that pink clip that I marked, I want to move it to this part because you're trying to reassemble. Look. Even though it's pink cut this.
09:15Okay. Look at the pink clip and I want you to move it and follow directions of the marker that I have, the cream colored marker. Also remove all the red markers as well.
09:29Put in that command. K. And I'm going to put on this screen now.
09:35That way, we can see what it looks like it's doing instead of moving the clip. It's doing is understanding where the clip is located and in relation to everything else, it's taking that clip, it's going to collapse the whole thing and it's going to move it to that spot.
09:52And there it is. It did this.
09:57It worked out pretty well. The unfortunate thing right here though is that during this process, it did did detach the clips from each other.
10:09So if we tried moving it again, do something else. Test it just in case.
10:14Now, I want you to move that pink clip all the way to the beginning of the footage.
10:23If it works or not. It's moving the clips again. See a pink clip all the way at the beginning.
10:32Is the clip clip because how we know is this white gap, but it did change the clip from pink back to its original color.
10:41There there we go. Never mind. It changed it back to pink afterwards.
10:45Have an idea of how this works. You can manipulate your clips however you like. And the next thing that you can do too is, let's say for example, you want to remove this silent space.
10:57You can either put a marker and so this is what we have done all the way up to this point. So right now our input so far is that we have our footage, our script, WhisperX.
11:07WhisperX is essentially a tool or an app that Claude uses in order to transcribe footage because a lot of these AI, they can't watch videos the same way that we as humans watch videos.
11:23They scan through each frame and dissect it, then they take your audio and transcribe into words. That is how they process videos at its core.
11:33Now, there's some AI like Gemini, it can realistically watch videos, but there's a certain limit to that, like a capacity limit. If you're more than two gigabytes, it can't watch it.
11:45So you have to trim it or cut it up and make it small. It's like the limit is two gigabytes and sixty minutes. So the first thing we did was we made sure or it made sure DaVinci Resolve was accessible on the computer.
11:59Then it set up a timeline, a 29.97 timeline. It imported all the footage.
12:06After that, it analyzed the footage. Now it takes the transcript and pretty much or it takes the video footage and transcribes it into words.
12:16It analyzes each word and it creates different files for each clip so that it can then align.
12:25So that is what it has done all the way up to this point in the background. If you don't have a script, I would say that you could still use this specific method to make your first cut.
12:37Just be very careful, literally hand hold it through the entire process.
12:43Have it transcribed all of your footage and then hand hold it through. Don't let Claude or any of your other AI agent make any necessary cuts without your permission at first until you can completely trust it. Question is, would you trust a junior editor or a brand new editor to your team with the first cut without giving them any directions at all?
13:07Just to that, maybe it depends. Me, personally, I would like to see how it does things first. I would rather be involved with the process and just train it.
13:17Just looking at the audio waves, it trimmed it pretty well. There was one section I noticed earlier where this one I know was kind of like a mistrimp, but that's okay.
13:28Like for the most part, could see it cut everything pretty well. You don't have to do any of that manually, which is the important thing.
13:36But you see right here, there's like a whole big chunk of space and right here as well.
13:43So what we're gonna do, just so that I can show you, is we're gonna ask the agent this. I want you to go through the timeline and mark all the spots with dead spaces with a red marker so that I can review and then I'll approve on whether you can cut out those dead spaces or not. We'll send that through.
14:02The idea here is to see how accurate it can be and if it even recognizes where there are dead spaces and where there's not. And the reason why you want to do this as well is so that whenever you tell it, hey, cut out the dead spaces, you can be sure that it cuts out the right dead spaces or, you know, once it marks it on here, we can go through and let's say for example, the space that it does is just too much, we can remove out that marker.
14:32Alright. So I had it go back and take a look. And so you can see it's not perfect.
14:39It's not perfect. I don't know what this cut is even. Like, the parts that it marked aren't necessarily the parts that we cared about.
14:46See, it could have marked right here. It's marking some pretty random spots. See, like, right here too.
14:52This both of these parts, unless Come on. Might be off.
14:59But you could see, this is what I mean by when I say these editors aren't perfect.
15:05You still have to go through and fix them yourself, especially if there's some dead air. Like, you can leave them, but I always make it a point to go back through and polish it. I mean, there's this beginning part too.
15:15That's too much dead space for introduction. Right? Or starting a video.
15:21And especially this part, this is the main portion that I would have liked it to cut out that it did not do. So a little bit bit disappointing, but this is why I say, still have your job and tell it at one second.
15:37It's in one second. There's a big gap, a gap.
15:41Do it yourself. You just do it yourself. It's a lot quicker than just telling it before you.
15:48See? Just like that, you're done. It's just some scenarios where doing things yourself so much quicker like here, there's a slight gap.
15:58I want to cut this whenever there's motion. Feels smoother.
16:05Just how it is. Right here, big gap at the end, we're gonna cut that, and we're gonna move this here.
16:14Now a chunk. What happened here? And that's the exact in a version of this into the actual Now, you wanna sweep through your entire video and clean it up yourself.
16:25Like, these are the small parts that really polish up and make the video look great. You just saved a whole bunch of time, like hours hours just making that first cut. So while your agents doing that first cut, you're not just sitting there.
16:40I'm usually off working on packaging, animations, sound, whatever the case is, but you're working in parallel now or side by side. It's much more complicated if the knot has been tied over itself many times.
16:52I'd rather just unravel it one knot at a time and know and understand what is currently happening. And that's kind of how I think about this. AI empowers you.
17:02It can only replace you if you're doing the exact same thing it does. Right now, it does the specialist work. The cuts, the assembly, the repetitive stuff.
17:12What it can't do yet is manage that whole thing. It doesn't know which pieces actually matter. The judgment is still yours and I think that's the skill worth building.
17:21At least that's how it's shaking out for me. If you want proof this works, I edit a real video start to finish with this setup in my last video. Go watch that.
17:30If you're setting this up, drop agent editing in the comments. Everything's linked down below in the description.
The Hook
The bait, then the rug-pull.
He was paying $350 a week — roughly $1,300 a month — to three freelance editors on a channel that was making zero dollars. So he set up Claude as his editor inside DaVinci Resolve instead, and this is the exact setup, start to finish.
Named ideas worth stealing.
05:31list
The agent-editing kickoff prompt
- Footage file path (or drag in the folder)
- Script / reference doc (Notion page or wherever the script lives)
- A scoped instruction (e.g. 'just the first two minutes')
- Explicit skill call ('use the brainstorm skill')
- Plan mode before execution
The five-part prompt template used to kick off a new Claude Code editing session: point it at the footage, the script, a scoped chunk of the video, the brainstorm skill, and plan mode before it touches anything.
Steal forany first-pass-cut prompt template for an AI editing agent
How they asked for the click.
VERBAL ASK
17:17next-video
“If you want proof this works, I edit a real video start to finish with this setup in my last video. Go watch that.”
Soft, single-ask CTA at the very end pointing to a companion video, followed by a comment-bait line ('drop agent editing in the comments') — no mid-roll pitch, no product sale.
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.
