Big Idea

The argument in one line.

ChatGPT's Codex, wired into DaVinci Resolve through an MCP server, can produce a usable first-pass edit for $20 a month if the editor writes down their own cutting rules first, since the workflow around the model matters more than which model is running it.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

A video editor or content creator who currently cuts every rough pass by hand and wants an AI agent to do that first pass instead.
Someone already comfortable giving an AI agent file access on their computer, or willing to learn safe approval settings before trying it.
A ChatGPT Plus subscriber, or someone considering it, who owns DaVinci Resolve Studio and wants to know if Codex is worth using for editing.
Someone who tried the Claude version of this workflow and wants a lower-cost alternative to compare it against.

SKIP IF…

You're looking for a one-click, no-setup tool. This requires installing a plugin, an MCP server, and defining your own cutting rules before it works.
You use the free version of DaVinci Resolve. The MCP server documents a workaround but the creator hasn't tested it.
You want finished output with graphics, motion text, or color. This produces a raw string-out only, no polish.

TL;DR

The full version, fast.

Andy Diep wires ChatGPT's Codex into DaVinci Resolve through an MCP server, then hands it raw footage plus a set of cutting rules: strip slates, false starts, wrong words, and dead space, and cut on movement so edits feel natural. Codex maps the footage, brainstorms an approach, and builds a full first-pass timeline inside Resolve on its own, about an hour across two runs for a cut he figures would take him one to two hours by hand. He reviews it, fixes one missed frame, then has Codex turn the whole process into a reusable skill. His verdict: Claude still wins on raw quality, but Codex on the $20 ChatGPT Plus plan is hard to beat on cost.

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 →

Chapters

Where the time goes.

00:00 – 00:40

01 · The finished cut, playing

Andy opens with the ChatGPT-built edit already running before explaining anything, then cuts to camera to preview what the video covers.

00:40 – 01:20

02 · Who this is for

He frames the video for creators and editors who are tired of cutting their own footage by hand.

01:20 – 02:01

03 · Why I tried ChatGPT after the Claude video

After his Claude + DaVinci Resolve video blew up, viewers asked for a ChatGPT version. He found Codex a genuinely solid contender, sometimes better than Claude.

02:01 – 02:33

04 · What you need before you start

The minimum requirement is the $20/month ChatGPT Plus plan for Codex access, plus paid DaVinci Resolve Studio. A free-version workaround exists but is untested.

02:33 – 02:47

05 · Step 1: download Codex

Codex lives inside the ChatGPT desktop app. He downloads and installs the app.

02:47 – 03:24

06 · What Codex actually is

Codex is OpenAI's coding agent that lives inside ChatGPT and can operate other apps on the computer, which is how it reaches into DaVinci Resolve.

03:24 – 04:18

07 · Step 2: install Superpowers

He installs the Superpowers plugin through Codex's Settings > Plugins > browse directory, and makes sure the brainstorming skill is switched on.

04:18 – 04:46

08 · Scoping the project folder

He creates a dedicated project folder so Codex only has file access inside that folder, not the whole computer.

04:46 – 05:36

09 · Step 3: the DaVinci Resolve MCP server

He installs Samuel Gursky's DaVinci Resolve MCP server from GitHub, the code that teaches the agent how to drive Resolve. A workaround exists for the free version but he hasn't tested it.

05:36 – 07:02

10 · Approval modes, and how people delete their own files

He recommends 'ask for approval' over full access for first-time setups, warning that users on full access have deleted their own files and blamed the AI for what was really a permissions mistake.

07:02 – 07:52

11 · The prompt

He hands Codex the raw footage folder and asks it to map every clip, then build a first pass that cuts wrong takes, dead space, and mistakes.

07:52 – 09:03

12 · The brainstorm, and the decisions I made

Under Codex's brainstorm mode he chooses 'best takes only' over a full string-out, then defines the cleanup rules: cut slates, action cuts, resets, false starts, wrong words, abandoned sentences, duplicates, and incomplete tails.

09:03 – 10:35

13 · How I actually cut a video

He explains his own manual process: transcription-based editing in Resolve, reading text instead of listening because it's faster, then a second pass to cut on movement so cuts feel natural.

10:35 – 11:40

14 · Resolve opens itself

Codex launches DaVinci Resolve on his main monitor via the MCP server and starts building the timeline live while he narrates, careful not to touch anything in case it crashes.

11:40 – 12:25

15 · The first thing it got wrong

He catches Codex leaving too much on some cuts and watches it self-verify by scanning back through its own work before he's satisfied.

12:25 – 13:38

16 · Cost, and why the model is not the bottleneck

The run stays under a single context compact at 217K tokens on OpenAI's cheapest accessible tier. His argument: the win isn't picking the priciest model, it's building the system around it.

13:38 – 14:48

17 · The run finishes, and I catch a miss

Two runs total 53:47 and 12:18, about an hour and five minutes for a cut he estimates would take him one to two hours by hand. Reviewing the timeline, he spots one place the agent missed a frame.

14:48 – 15:34

18 · Turning the session into a reusable skill

Once he's happy with a first pass, he tells Codex to package everything it just did into a reusable first-pass skill, so future footage in the same format can be run with a single prompt.

15:34 – 16:34

19 · Which one I would actually use

His verdict: pick Claude/Fable for raw quality since it handles more across the board, even though it costs more. Pick ChatGPT/Codex on the $20 plan for value, since it also generates images and can write scripts.

16:34 – 17:02

20 · Close

He closes with a recap card of the tools and links he used and points to his earlier Claude-based version of this same workflow.

Atomic Insights

Lines worth screenshotting.

A $20 ChatGPT Plus plan is the entire minimum spend needed to run Codex against DaVinci Resolve Studio for a first-pass edit.
Codex reaches into DaVinci Resolve through an MCP server, the same kind of connector that lets an AI agent operate other apps on your computer.
Scoping the agent's file access to one dedicated project folder means a runaway command can only touch footage inside that folder, nothing else on the computer.
People who give an AI agent full access without understanding what it changes have deleted their own files and blamed the AI for what was really a permissions mistake.
The cutting rules that matter most are explicit: cut slates, action cuts, resets, false starts, wrong words, abandoned sentences, duplicate deliveries, and incomplete tails.
Reading a transcript is faster than listening to raw audio when deciding what to cut, which is why transcription-based editing beats scrubbing by ear.
Cutting on the moment a person starts a physical motion, not on the words, is what makes an edit feel natural instead of mechanical.
A two-run first pass took 53 minutes and 12 minutes, about an hour and five minutes total, for a cut estimated at one to two hours by hand.
The agent ran the entire session on OpenAI's cheapest accessible model tier and never hit a single context compact at 217,000 tokens.
Don't min-max your model, min-max your workflow: the system built around the AI matters more than which model is driving it.
Once a first pass works, asking the agent to package the session into a reusable skill turns a one-off setup into a repeatable command for future footage.
For raw coding and editing quality, Claude's Fable model still outperforms Codex across the board, according to someone who has used both extensively.
Codex on the $20 ChatGPT Plus plan adds a built-in image generator and can write scripts, on top of driving the editor directly.
The model isn't the asset. The system built around it, the rules, the folder scoping, the approval settings, is what actually gets reused.

Takeaway

How to Run an AI First-Pass Edit

WHAT TO LEARN

The real skill isn't picking the best AI model, it's writing explicit cutting rules and safe access boundaries once, then reusing that system on every future edit.

04What you need before you start

The entry cost for this workflow is a $20/month ChatGPT Plus subscription plus a DaVinci Resolve Studio license, since the free version isn't officially supported.
Check the minimum plan tier before you commit. Automation like this is gated behind mid-tier subscriptions, not the free version of the app.

06What Codex actually is

An AI coding agent that can operate other applications on your computer is a different category of tool than a chatbot. It reaches past its own window into whatever software you point it at.
Understanding that a coding agent is what's driving your editor, not a purpose-built editing AI, sets the right expectations for what it will and won't do well.

08Scoping the project folder

Before giving an agent file access, create one dedicated folder for it to work in rather than pointing it at your whole computer or drive.
Scoping access to a single folder is the cheapest insurance against an agent mistake: the blast radius of anything that goes wrong is limited to that folder.

10Approval modes, and how people delete their own files

Start any new agent workflow on 'ask for approval' rather than full access, until you've watched it operate enough times to trust its judgment.
Most horror stories about an AI 'deleting everything' trace back to a user granting full access before understanding what the agent could touch, not a rogue model.

11The prompt

A first-pass prompt only needs two things: where the footage lives, and what counts as a mistake to cut. Everything else can be figured out by the agent.
Letting the agent map the footage first, before asking it to make decisions, gives it the same grounding a human editor gets from watching all the raw footage once.

12The brainstorm, and the decisions I made

Writing down cutting rules as an explicit list, slates, false starts, wrong words, dead space, duplicate takes, incomplete tails, turns a vague 'clean this up' request into something an agent can execute consistently.
Choosing 'best takes only' over a full string-out at the planning stage avoids reviewing every raw take later. The decision gets made once, up front, instead of during cleanup.
A brainstorm or planning phase before execution is what keeps an agent from making irreversible choices on your behalf. Skipping it is what makes results feel random.

13How I actually cut a video

Transcription-based editing, cutting by reading text instead of listening to audio, is measurably faster because reading beats listening for spotting what to keep.
A two-pass method, one pass to place rough cuts by transcript, a second pass to listen and polish take quality, catches problems a single pass misses.

15The first thing it got wrong

Even a well-scoped agent run still needs a human review pass. Errors like a monitor left visible too long are the kind only a second set of eyes catches.
Watching an agent self-correct by scanning back through its own output is a sign the tool is checking its own work, not a substitute for you checking it too.

16Cost, and why the model is not the bottleneck

The cheapest available model tier finished a real, complex task without hitting a single context compact, evidence that raw model horsepower matters less than a well-defined workflow.
Min-max your workflow, not your model: the payoff comes from perfecting the rules and steps you feed an agent, not from always buying the most expensive tier.

17The run finishes, and I catch a miss

Two runs totaling about an hour and five minutes replaced a task estimated at one to two hours by hand, a real but not dramatic time save on a first pass.
The honest ROI case for agent-driven editing is a modest time reduction on tedious work, not a claim that it replaces editorial judgment.

18Turning the session into a reusable skill

Once a workflow works, the next move is asking the agent to convert that exact session into a named, reusable skill, so the next similar job runs in one prompt instead of a full walkthrough.
Turning a one-off automation into a repeatable command is what separates a demo from something you'll actually use again.

19Which one I would actually use

The choice between AI tools comes down to a real tradeoff: pick the higher-quality, more expensive model when output quality is what you're being paid for, pick the cheaper model when you're optimizing for cost.
A model that also handles adjacent tasks, like image generation or script writing, can be worth choosing over a strictly higher-quality model if it replaces multiple subscriptions.

Glossary

Terms worth knowing.

Codex: OpenAI's coding agent that lives inside the ChatGPT app and can operate other applications on a computer, which is how it reaches into a video editor.
MCP server: A connector program that teaches an AI agent how to control a specific piece of software, in this case DaVinci Resolve, by exposing its functions in a way the agent can call.
Superpowers plugin: A Codex plugin that adds structured workflows like a brainstorming mode, installed from Codex's Settings > Plugins menu.
First pass: A rough, unpolished cut that removes obvious mistakes and dead space but has no zooms, graphics, or text animation, meant as a starting point for a real edit.
String-out: A basic assembly of clips in order with no creative cutting, the simplest possible first-pass edit.
Approval mode: A Codex setting that controls whether the agent asks for permission before taking an action, versus acting freely with full access.
Brainstorm mode: A Superpowers workflow where the agent proposes and discusses an approach with the user before executing it, rather than jumping straight to work.

Resources

Things they pointed at.

00:00toolDaVinci Resolve MCP server (Samuel Gursky) ↗

03:30toolSuperpowers plugin

00:20productChatGPT Plus / Codex

00:00toolDaVinci Resolve Studio

16:31linkThe Claude + DaVinci Resolve version (Andy's earlier video) ↗

Quotables

Lines you could clip.

00:34

“If you're a creator or an editor tired of sitting there cutting your own footage, this is just for you.”

direct audience callout, no context needed→ TikTok hook↗ Tweet quote

13:21

“Don't try to min max your model, but min max your workflow.”

tight systems-thinking line, quotable on its own→ newsletter pull-quote↗ Tweet quote

16:22

“The asset is your system. The model is just there to connect everything together.”

closing thesis, stands alone with zero setup→ IG reel cold open↗ Tweet quote

The Script

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

So what you're seeing on screen right now is ChatGPT. It transcribed.

It organized all my footage, and then it cut everything clip to clip so that there is no dead space in between and it cut it very well as like shockingly, it rarely cut on the parts where I'm still speaking.

Now, I'm gonna show you in this video the parts that ChadGPT did really well, the parts where it kind of messed up, and how to set everything up starting with setting all of this up from nothing.

If you're a creator or an editor tired of sitting there cutting your own footage, this is just for you. Cutting has always been one of those brainless activities for me. You just cut, cut, cut again and again.

And after doing it for the last four years professionally, I have it down to a science. I know what to cut, what to keep, and when to make the cut. There are a series of steps involved in it, and for me, it's not something that requires a lot of creativity.

It's just step one, step two, step three every time. So I taught Cachapiti these steps.

It runs on its own for about an hour. You go to lunch, come back to a finished cut, look it over, and get on with the part you actually like. Okay.

So in my last video, the Claude plus DaVinci Resolve video, it blew up and there was a lot of comments in the video that was asking, what about ChatGPT?

I would like to do this with ChatGPT. And so I'm gonna give you guys what you have asked for. I thought that I really did not think that ChatGPT would perform as well or even better than Claude, but surprise surprise, in one of my previous videos, I tested it out and ChatGPT is actually a very solid contender against Claude, and there were instances in which it was even better than Claude.

So I'm gonna show you guys how to set it up in this video. And do note that the smallest thing or the smallest plan that you can do this with is the $20 codex plan.

So let's go to their plan real quick. So there are a few different tier plans. You need to have at least the plus plan.

The plus is the only way for you to connect or to use codecs in order to connect ChatGPT over to DaVinci Resolve. And you will need the DaVinci Resolve Studio.

If you don't have those two, this will not be able to work. You want to go to OpenAI. I'll also have a PDF down in the description and pinned comment that you can follow or just literally throw into your ChatGPT or even Claude.

Codex is OpenAI's coding agent. It lives inside the ChatGPT app, and it can operate on other apps on your computer, which is how it connects to DaVinci Resolve.

So we are gonna go over to codecs right here and download.

You're gonna hit download for Mac OS. You're gonna install that. It'll walk you through it.

You'll have this, drag this. If you're on Mac, drag it in, or if you're on Windows, just double click and install it. It's still called Codex, but the whole app itself is called ChatGPT.

And so you wanna go up to here where it says ChatGPT, this little arrow, and click on Codex. Once you're in codex, we are going to install superpowers.

So in order to install superpowers, you're gonna click on your name at the bottom left here, go into settings, and you'll look for plugins.

Right here under integrations, you'll click on plugins. Once you click on the plugins page, you'll click up here which says browse directory, click on that and search up superpowers.

Right here, I already have installed so I don't have to install it, but you can click on it and right here should be a little install or like to check it on, you'll click that. Down here you can see some of the skills that it has. Make sure that brainstorming is on.

Brainstorming is the thing that we need the most from here.

So once you have installed superpowers, you'll be on this page now, and it should be relatively blank. You'll click on choose project, new project, and let's go ahead and create a folder inside of your directory.

Go into this little house icon that you have and create AI assistant editor.

Basically do all the things that you can, but only inside of this folder. So as long as you don't give it access to anything else inside of your computer, this folder is the only folder that has access to files. Now we're gonna go into GitHub.

Now GitHub, it is basically code that teaches our agent how to use DaVinci Resolve.

I'm not gonna get too much into this, but how it works is that you essentially go ahead and copy this MPX DaVinci Resolve MCP setup.

So it looks like with DaVinci this DaVinci Resolve MCP, there is a workaround for the free edition, for the free version of DaVinci Resolve.

But I am not well versed in that, so if that's something that you want to use, you may have to figure that out on your own with the help of ChatGePety or Codex. So let's turn on slash plan and then you can do slash brainstorm.

One key thing that I would recommend that you do that is different from what I do is full access. If you're just doing this for the very first time, which I assume most of you are doing, I would click on ask for approval until you fully understand what Codex does on each step.

It is not a stretch of truth to say that some people who do not know what they're doing, they have put full access, they have done a full send, and they have deleted a lot of them, then they go on x and post like, Claude did this to me when it was really a user error. So be very careful with putting it on full access if you do not know what you are doing yet.

So now that we're in here, we are going to take this footage file that I already have that I placed inside my AI assistant editor folder. We'll go ahead and copy it and we'll give it over to CHEJBT.

And so whenever it does things like this, whenever it does any thinking, any planning, I would always recommend you use the highest setting. Once this planning mode is done, we can swap this off back to 5.6, and this maximizes your efficiency in terms of token spend, like we're putting the best mind to thinking, and we're also now switching over to the best hands to execute.

Hey, chat. I want you to go ahead and map out all the footage that we have inside this folder, then let's turn it into a first pass where we cut out any spots that are wrong, that say the wrong word, any parts that we mess up, and any dead spaces.

So you can see right now it's searching the files, it's mapping out all the footage that's inside of our folder right now.

If there's no transcript, what it will do is it will take each video file and start transcribing it by itself. And we have a passable first pass, at least that's that's why I call it a first pass.

So you see, we can have the full story, a clean string. Yeah, we are going to do the clean string out.

So we have two timelines, all usable takes or best takes only. So we are going to do best takes only. So it has come up with three workable approach.

Number one, rebuild from the audited map and verify against these local file. This is my recommendation. Relink only to this folder preserved source.

So we're gonna go with step one or I mean, option one. And once you have the skill, let's see, cleanup rules.

And this is the most important part right here that saves editors a lot of time. It's the cleanups section. So start from the audited, select slash reject time alone, listen to every range, remove slates, action cuts, resets, false starts, wrong words, abandoned sentences, duplicate deliveries, and incomplete tails.

So this is where the plan, this option that we had checked earlier kicks in.

This whole time, we were underneath the brainstorming. This whole process is brainstorming. Okay?

This isn't like AI will just do everything for you. This is not a click one button and AI does it. That isn't what this is at all.

So once I do that, how I would cut and edit a video is I do it by transcription base. What that means is I go into DaVinci Resolve and I have DaVinci Resolve transcribe every single clip.

I highlight the text, I do insert, and that is how I basically cut the video and read rather than listen because listening, I found, is significantly slower than reading it.

So once that is all done, I will go back and I will start polishing every single cut. So in the next pass, I go through, I listen for the best takes, and I remove all the takes that I did not like or has been repeated or has been messed up.

Whichever the case is, I have the polished pass that I do later on where I'm very precise with where I cut enough. The the goal here is to make a cut of when the person starts a motion, like for example, right here.

If we go back right when I move my hands, like the moment I start moving my hands, that's when I want to make the cut because we're making a cut on a movement, so it looks more natural. That is the gist of things.

While this video is still doing some work, honestly, if I were to be cutting up this video, it would probably take me at least one to two hours.

Oh, you guys won't see it on your side but I'll move it in. DaVinci Resolve is currently being opened on my computer, on my main screen, not my laptop.

So once that opens up, I'm actually going to drag it over. Oh, but you can see right here, it says, oh yeah, yeah, right there, see? It's opening up.

You can see that Codex is now looking at this.

I can't move this. If I move this or play with it too much, it might just accidentally crash.

So I will move it. But yeah, you can see that right here Codex had is just operating DaVinci Resolve because of the MCP server from earlier.

Is designing the timeline now and the resolve project build.

And there we go. It has just launched DaVinci Resolve on my main.

So the issue the persistent issue that we're seeing here is that these monitor at least x is leaving too much So what is happening is I think ChadGBT is watching through the whole entire video. It's like scanning through. It is Let put up my mind.

Like, it is self verifying and making sure I have never used 5.6 Luna like this until I saw that x.

This is actually still so crazy. It it's still going, guys.

It's still going. The reason why I think it's so crazy is because we are literally on 5.6 Luna extra high. 5.6 Luna, you have to understand, this is the cheapest model that OpenAI has that's accessible.

And more recently as well, they have completely reduced the price of 5.6 Luna in order to compete with a lot of the Chinese models. And only 217 k tokens, we haven't ran a single compact yet either, which is huge.

Running this while you are reviewing over your b roll or you're doing your b roll cuts, Like, is literally just verifying for you your a roll pass. A roll pass is probably one of the most boring. You can always reshape that later as well.

I don't know. It's just honestly, the truth is the sweet spot is somewhere in the middle.

There's a mixture of human involvement, and there's a mixture of agent involvement. It isn't one side or the other, but one side or the other is just so my recommendation is don't try to min max your model, but min max your workflow.

As long as you perfect your workflow in the systems that you run through your agents, it doesn't matter what model you're using. As you can see, I'm using the lowest tier model, saving a whole bunch of tokens as well. Okay.

Well, it literally has just finished after I paused it like ten seconds ago. Look, sometimes SharedHBT will ask you for access like this.

I always just click don't allow. Fifty three minutes forty seven seconds for that run, and this one was twelve minutes and eighteen seconds.

So altogether, we're looking at about one hour and five minutes where see, I think there was just, like, since this wasn't one frame, it didn't catch it. And literally, let's review through this timeline now.

See, I think there was just, like, since this wasn't one frame, it didn't catch it.

But I guess you don't I think it followed my rule.

Now the last step you will do is once you are done with all this and you're like very happy with the first pass that it did, literally prompt it. Now I want you to create what's called the first pass skill based off of everything that we just did. I already have a first pass skill, so I'm not going to press enter, but if you don't have one already, go ahead and press enter, and the next time you have any footage that is similar in a format like the talking head, you just say, hey, run first pass skill, and it will literally just do the exact same steps that we just did here, like, and that's all.

Second time of running faster, this time we didn't even have it on fast either. If you put this on fast, it would have saved a lot more time as well. We We probably could have cut this down to 1.5 times.

It's half the time about Now I know a lot of you are gonna ask which one is better. If quality is what you care about the most and you do work beyond just content, stick with Claude. Fable five, Claude's top model, is the best coding model I have used and tested, and it handles more across the board.

The motion graphics and the text animations, I still build those in Claude, but it's also one of the most expensive. Codecs on the 20 Chachi PT plan is really hard to beat.

It has an image generator built in. It can drive Resolve as your editor, and it can write scripts if you want. It is not going to match Fable's quality.

I'm just putting that out there. Don't commit to one model and believe it is the end all, be all. The asset is your system.

The model is just there to connect everything together. If you wanna see this same workflow proven end to end through Claude, that video is right here.

But, yeah, that is the full walkthrough of how I set up DaVinci Resolve with Codecs and then edit with it. It's very simple, very straightforward, at least for me.

If you guys have any questions on how to do this, like you're still confused or something or you want something a little bit extra, leave a comment down below this video and I'll answer it as I see it. Alright.

That's my dog.

The Hook

The bait, then the rug-pull.

Before he explains anything, Andy Diep shows the finished product: a DaVinci Resolve timeline that ChatGPT's Codex cut on its own, clip to clip, with almost no dead space in between.

Frameworks