A paid-social creator feeds Claude Code raw footage and a script, gets back a finished Facebook ad with no editor and no timeline, then shows how feedback turns into a reusable editing skill.
Posted
2 days ago
Duration
Format
Demo
educational
Views
3.1K
58 likes
57 · 43
Big Idea
The argument in one line.
Feeding raw footage and a script directly to Claude Code produces a workable first-draft ad with no editor or timeline software, and writing explicit feedback into a reusable skill is what turns that first draft into a repeatable, improving editing workflow.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
Someone running paid social ads who currently pays an editor for every cut, variation, and revision round.
A creative director or ops lead curious whether an AI coding agent can realistically replace parts of an editing pipeline.
Anyone already using Claude Code or Codex for other work who wants a concrete non-coding use case.
A solo marketer or small team without a dedicated video editor who needs ad variations produced quickly.
SKIP IF…
You need frame-accurate color grading, complex motion graphics, or anything beyond straightforward cut-and-caption editing.
You're looking for a literal prompt template; the video shows the workflow and the result, not a copy-paste script.
TL;DR
The full version, fast.
A paid-social creator fed Claude Code, running on Opus 5.5, a Google Drive folder of raw footage and a script from Notion and got back a finished Facebook ad with no editor, no Premiere Pro, and no manual timeline. The agent itself downloads the footage and writes the edit; it isn't controlling a browser or driving DaVinci Resolve's own automation. The first draft isn't perfect, but natural-language feedback fixes it inside the same chat, and writing that feedback down as explicit rules turns it into a reusable 'video editing skill' that produces stronger first drafts every time it runs. The bigger claim: as agents take over execution, documenting the reasoning behind creative decisions becomes the real differentiator.
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.
Cold open: Claude Code turned raw footage and a script into a finished Facebook ad with no editor, no Premiere Pro, no timeline.
00:14 – 03:32
02 · The One-Shot Edit
The one-shot prompt (raw footage link + script, nothing else), the first finished ad it produced, giving plain-English feedback, and how the agent technically edits the footage itself rather than driving a browser or DaVinci.
03:32 – 04:19
03 · Edit Variations
Asking the agent to produce four variations of an already-approved ad, rendered unattended in about 20 minutes.
04:19 – 05:32
04 · Making a Skill
Converting specific feedback (B-roll placement rules) into a saved, reusable video editing skill.
05:32 – 07:27
05 · Document Learnings
The case for having editors, strategists, and creator sourcers narrate the reasoning behind their decisions so it can feed future skills.
07:27 – 08:56
06 · Reviewing Edits
Reviewing six AI-made edits at once and giving voice-dictated feedback; a second ad (cat scratcher product) after multiple feedback rounds.
08:56 – 10:53
07 · Creating Different Skills
Building narrow skills for QA/spellcheck, footage-to-brief matching, and distinct visual styles; framing today's output as the floor, not the ceiling.
10:53 – 11:22
08 · Summary
Closing argument to start working agentically or fall behind, plus a question to viewers and a like/comment/sub ask.
Atomic Insights
Lines worth screenshotting.
A finished Facebook ad came out of Claude Code from nothing but a Google Drive folder of raw footage and a Notion script, with no editor and no timeline software touched.
The agent downloads and edits the actual footage itself; it is not remote-controlling a browser or driving DaVinci Resolve's automation layer in the background.
Running the highest reasoning effort on a video edit barely dented a weekly token budget, even across multiple edits run in the same week.
A roughly 20-minute render time is irrelevant when it happens unattended in the background while other work continues.
Asking for variations of an already-approved edit is a single instruction, and the agent produced all four without additional direction.
Written, specific editing rules, like only cutting to B-roll at the start or end of a line, compound into a reusable skill instead of one-off instructions.
The single biggest lever for improving AI-driven creative output is getting editors, strategists, and creator sourcers to narrate why they made each decision, not just what they decided.
A reviewer can go from making one edit at a time to reviewing six finished AI-made edits in one sitting and giving feedback by voice.
Multiple rounds of natural-language feedback compound noticeably: an ad after two or three rounds reads materially tighter than its first V1 draft.
Narrow, single-purpose skills, one for spellcheck and brand QA, one for matching footage to a brief, one per visual style, outperform one general-purpose editing instruction.
Takeaway
Editing becomes a skill you write down, not software you operate.
WHAT TO LEARN
The jump from a rough one-shot AI edit to a reliable, repeatable one isn't a better prompt, it's writing feedback down as explicit rules and turning them into a saved skill.
02The One-Shot Edit
A full creative brief was nothing more than a Google Drive folder of raw footage and a script in Notion, with no manual editing before the first draft existed.
The agent itself downloads and edits the footage; it isn't remote-controlling a browser or driving DaVinci Resolve's own automation, a different approach than most AI-editing demos online.
The first one-shot draft wasn't perfect, but natural-language notes like 'change this clip at 32 seconds' or 'tighten the cuts' fixed it inside the same chat, no timeline required.
Running the highest reasoning effort on a video edit used far less of the weekly token budget than expected, even across multiple edits in one week.
03Edit Variations
Once one edit is approved, asking for variations of it is a single instruction, and all four rendered without further direction from the creator.
The roughly 20-minute render time didn't cost any attention because it ran in the background while other work continued.
04Making a Skill
Specific written rules, like only cutting to B-roll at the start or end of a line and only when it's genuinely relevant footage, replace vague one-off notes.
The exact trigger matters: after giving feedback, telling the agent to 'turn this into a video editing skill' is what makes the rules persist across future edits.
A workflow that starts with no context produces inconsistent V1s; the skill is what compounds today's feedback into tomorrow's better starting point.
05Document Learnings
The core shift is from making decisions to narrating decisions: an editor should explain why they made a cut, not just make it, so there's something for the agent to learn from.
The same habit generalizes past editing: a strategist explaining why a script is structured one way, or a sourcer explaining why a creator fits the brand, feeds the same kind of skill.
Recording those explanations and turning the transcripts into skill instructions is framed as the real differentiator between top and average creative teams, not the tool itself.
06Reviewing Edits
Once edits run unattended, the job shifts from producing one cut at a time to reviewing a batch of finished drafts, six at once here, and giving feedback by voice.
A second or third round of feedback on the same ad produced a result judged ready to run in the ad account without any editor involvement.
07Creating Different Skills
Narrow, single-purpose skills, one for spellcheck and brand QA, one for cross-checking footage against a brief, outperform one general instruction trying to do everything.
Splitting skills by visual style, an 'ugly ad' style versus a professional high-production style versus motion graphics, keeps each skill's instructions specific enough to actually follow.
The framing for where this goes: today's output is described as the floor, not the ceiling, with the gap to professional-suite editing expected to keep closing.
Glossary
Terms worth knowing.
Opus 5.5
The Claude model version used for the video edits in this walkthrough, run through Claude Code rather than a chat interface.
Claude Code
Anthropic's coding-agent tool, used here to perform video editing tasks, downloading footage and making cuts, rather than writing software.
Video editing skill
A set of written, reusable editing rules saved inside Claude Code so the agent applies the same standards automatically on every future edit.
DaVinci MCP
A Model Context Protocol integration some creators use to let an AI agent drive DaVinci Resolve's own editing interface, a different approach from the direct-footage editing shown here.
One-shot edit
A finished edit produced from a single prompt with no follow-up feedback, used here as the baseline 'V1' quality test.
Resources
Things they pointed at.
00:38toolGoogle Drive
00:38toolNotion
02:46toolDaVinci Resolve
02:46toolAdobe Premiere Pro
08:27toolChatGPT
07:30toolWhisper Flow
01:20productiRestore Elite + Lumina face mask
07:44productCatasaurus Peekaboo Cat Cave
Quotables
Lines you could clip.
00:00
“I have clawed a folder of raw creative footage and a script and it turned it into a finished Facebook ad. No editor, no Premiere Pro and no timeline.”
States the entire claim in two sentences with zero setup needed.→ TikTok hook↗ Tweet quote
10:07
“This is the worst the tool's ever going to be. The editing is only going to get better.”
Short, quotable framing line that works as a standalone caption.→ newsletter pull-quote↗ Tweet quote
05:29
“we are just the quality control for these AI agents”
Reframes the human role in one clean sentence.→ IG reel cold open↗ Tweet quote
10:37
“If you do not start working agentically then you are going to fall behind from the brands who have strategies who have editors who are working like this”
Direct urgency statement, good closing-argument clip.→ TikTok hook↗ 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.
17px
metaphorstory
I have clawed a folder of raw creative footage and a script and it turned it into a finished Facebook ad. No editor, no Premiere Pro and no timeline. It literally gave me the finished ad in clawed code.
Today I'm going to show you the actual edit alongside the exact process and how you can do this for your brand. So let's have a look at this because when I started playing with Opus 5 .5 for videos and I was in shock. I couldn't believe the output it was giving me.
And it's now to the point where we're doing a lot of our edits inside of Opus 5 .5. So the prompt here that I entered is I'm testing how good your video editing capability is. Here's the link to the raw files.
Literally gave it the raw files in a Google Drive folder and I gave it the script in Notion and that is it. This is one shot. As you can see, I have not dressed this up.
I have not edited this at all. I'm going to show you the exact edit that it gave me back one shot. And it's not perfect, but it is really impressive and much better than I thought I was going to get back as a V1 of an edit from Cloud Code.
Okay, if you're thinking about red light therapy for your hair and your skin, this makes so much sense. Because the thing with any kind of hair or skincare routine is you actually have to stick with it. And I don't need another hour -long routine at the end of my day.
So this is the iRestore Elite. You literally just put it on for 12 minutes a day and it's completely hands -free. It has 500 medical -grade lasers and LEDs covering your entire scalp.
And this is their Lumina face mask. It has red light, infrared light, and blue light, so you basically got different modes depending on what you want to target with your skin. And then I realized, why am I doing these things separately?
I could do my hair, my skin, watch TV, read, whatever, all at the same time. I call it my stacking routine. Instead of adding more things to my day, I'm just stacking them onto the time I already have.
And iRestore actually sells the Elite and the Illumina together as a bundle. So if you want both, you can save versus buying them separately. Honestly, if I could work on my hair and my skin literally while watching TV, why would I not?
There you go. I mean, again, not perfect. There are parts of that edit that are a little bit jolty and that you'd want to change.
But because it's in Cloud Code, you can literally just give it feedback in natural language and say, change this clip at 32 seconds. I can give it whatever feedback I want inside of here. I can say, change this clip.
I could say, tighten the cuts inside of here. And whatever you want to do, you could get this edit to the point where it's finished. inside of Chord Code without ever needing to be inside of Premiere Pro, without ever needing to be inside of DaVinci.
Let me explain what's actually happening here because a lot of people have seen people online editing with the DaVinci MCP or editing inside of Astra. This is done by the agent. So Chord is actually going to download this footage and the agent is writing the edit onto it and then giving it back to you.
It's not controlling my browser. It's not actually editing in DaVinci in the backend and give it back to me.
It is doing it inside of Cloud Code itself. It is doing it through Opus 5 .5. I ran this on Opus 5 .5 high, although I've been running a lot of my edits on medium.
To be fair, I've been realizing regardless of what effort I'm putting on. Opus 5 .5 does not take up a lot of tokens. I would have thought to go and edit this, it would take a huge amount of tokens and it would just burn through my weekly usage.
It's not doing that. I have had scenarios last week where I've been editing multiple videos at a time. It's barely taken up any of my token usage.
Now, granted, this did take about 20 minutes to run. But it was all in the background and I was doing other tasks while I was going to edit my videos for me without an editor, without anything needed from me, except for just me just saying, here's the raw footage and here is the script. And you can actually take it one step further.
If you have your footage and your brief linked and hooked up through connectors into Cloud Code or Codex, you can get it to do variations as well. I had this ad that we were producing for a supplement brand where... it went and did all four of the variations for me.
I said, here is the original ad that you made for me, go make the variations. And it literally went and made all four of the variations for me. And now I've gone from, you know, needing to get this video edited to having it edited and having all four variations just like that.
Now here's where it gets really interesting. These video edits are not perfect, especially if you just put in no context, like I have done here. And you just say, literally, here is the footage and here is the brief.
They're not going to come out perfect every single time. I thought this was okay, but there are still parts of this that I would improve in terms of an editing perspective. What you want to do is you want to turn this into a skill and give it feedback so it gets better over time.
So I could say here, for example, when you cut to B -roll, you must make sure that the B -roll that you're cutting to is relevant footage. I could say, when... you are making cuts to B -roll.
You should only do this at the start or the end of a line. You should never cut to B -roll midline or whatever you want to put in here. And you just put down all of your rules for how to go about conducting a good edit.
And then at the end of it, you just say, now turn this into a video editing skill. So the next time I run this skill and I give you an edit, you know to abide by these rules all the time. And that's going to create you a video editing skill that you can just say every time you upload a piece of footage or you give it the link to the folder like I have done here, you can just say, run the video editing skill and it will go and apply all of the feedback that you've given.
And then the more you use this, the more feedback you give it, the stronger it gets over time. And you'll get to a point pretty quickly where like it is putting out very, very good V1 edits that sometimes you don't even need to give any feedback to. if you have given the right context to the agent.
And that's why I think in a world where AI agents are doing all the work and the humans are just the orchestrators, we are just the quality control for these AI agents, one of the most important things that you can be doing as an organization is having everyone on your creative team explain why they are doing what they are doing as they are doing it.
So let me give you an example of what I mean by that. If you have editors that are currently editing in Premiere Pro or in DaVinci or whatever they're editing on, you should be asking them to, as they're editing, explain why are they making a cut?
Why are they making that cut there? Why are they deciding to paste the video like this? Why are they choosing this piece of B -roll rather than that piece of B -roll?
If you can get them to document all their learnings, record those calls, and then use the transcripts to inject into a skill like this, then you have a skill that is really, really powerful. And your editors get to have Claude Code or Codex be the editors, and they can just be the orchestrators who are reviewing 10 different edits at the same time rather than making one edit at the same time.
I couldn't be a big enough opponent for this, whether it's editors or strategists or creator sources or whatever. If you are a strategist, you should be recording yourself doing strategy as you're doing strategy and explaining why you're writing a script like this, why the copy flows from this line to this line, why the script is structured like this instead of this, why you've chosen this hook, why you've chosen this idea over this one.
If you're a creator sourcer, what do you love about this creator? Why are they going to be a great fit for your brand? And taking all of this context and injecting it into skills, injecting it into your company brain, because that is what is going to separate the top companies at Ad Creative and in general from the average and mediocre brands when it comes to paid social, when it comes to AI uses in general.
Because as you guys know, context is everything. And the better context that you have about your company and the travel knowledge and the domain context of how to complete a task, the better off your outputs are going to be. And once you have that, a world like this is very possible where you can just be coming into Cloud Code and reviewing the edits that it's made for you.
I mean, I was working on some edits the other day and literally I came into Cloud Code and reviewed six edits at once. And I was just able to come in here, watch the videos, whisper flow, voice dictate some feedback, and then it would go back to work on the edits. And I'll play another one through here.
And this one, I think, went through two or three rounds of feedback. Let's see how it came out. I had no idea that this small change would save me a $400 vet bill.
I used to spend multiple hours a day, hundreds of dollars, constantly cleaning up and replacing these cheap cat scratchers that my cat would tear up everywhere. When my cat doesn't use the cheap cardboard things, she uses the couch, the bed, anything she can get her nails into. I don't know about your cat, but my cat's nails grow almost faster than mine.
I stumbled upon the Catasaurus Peekaboo Cat Cake. She can hide in it, she can scratch it, and she can even sleep in it if she wants to. What I love about it too is that there's no mess when she scratches it.
The cat cave is made with a specialized felt so that it actually helps her nails from getting overgrown. Even the vet noticed. She was like, what are you doing to her nails?
The peekaboo cave ended up solving two of our problems at once. One, her nails, they're healthy and she gets to hide. And there you go.
I think that is pretty much good to go. I mean, there are a couple of points there, maybe with the text overlay that need adjusting, but like that is good to run inside the ad account. And I have not had an editor work on that at all.
That is all clawed code. uh agents working to edit the videos for me which is just crazy like it's crazy how much of the process you can now do with ai i uh have uh built skills inside of my core code and codex that go and review the creator footage so it will go into the google drive and say look at the footage here and cross check against what is in the brief and then contact the creator if they are missing any of the clips in here.
We've built skills inside of here that will spell check and QA the final ads. Skills that will make different variations of ads. Skills that will edit for different styles of ads.
And that's the next iteration of this, I guess, is to make skills for every type of edit that you make. So you might have one that edits for ugly ads. You might have another skill that edits for professional high production ads.
You might have another skill that edits for, you know, motion animated graphics or whatever it be. But like you can have specific skills for those specific edits and then you can just say, oh, apply the air animation edit skill on this ad and it goes and edits like an air animation ad because you can put just the context of those ads in there so that the instructions are more clear and specific to that format.
But you really can get to a point where, unless it's a really complex edit, then you could be able to do most of your editing inside of Cloud Code or Codex. And the exciting thing is, This is the worst the tool's ever going to be.
The editing is only going to get better. This is like the V1 of a code and codex editing. So what's it going to be like in a month time, three months time, six months time?
It's going to be at the point where you don't need to do any editing. You don't need to be in Premiere Pro. You don't need to be in DaVinci.
You can literally just voice dictate in natural language what you want the agent to do and it will go and do it for you. This is a drastically different way of working. The way a creative team is operating is changing in real time.
If you do not, uh start working agentically then you are going to fall behind from the brands who have strategies who have editors who are working like this and having called code go and edit 10 videos for them at the same time because then you can just get so much more done you can put out so much more ads you can test machine quickly and you can find out what works for your brand so let me know have you tried opus 5 .5 for editing yet uh do you like it have you started doing a lot of editing inside of there do you not like it i'm curious to hear if you found this interesting drop a like comment what you guys want me to make more content about i do enjoy making these ai add creative style videos.
Because like I said, this stuff is changing in real time and it has been, a creative team is literally being redefined in real time, which I really find fascinating. So if you enjoyed it, like, comment, sub, all that good stuff. And I will see you guys in the next one.
The Hook
The bait, then the rug-pull.
Claude Code turned a folder of raw clips and a script into a finished Facebook ad with nobody opening an editing app. This breakdown walks through that one-shot edit, a second ad that went through several rounds of feedback, and how writing that feedback down turns into a reusable editing skill.
Frameworks
Named ideas worth stealing.
04:11model
Feedback-to-Skill Editing Loop
Give the agent raw footage and a script
Review the one-shot V1 edit
Give specific natural-language feedback on what's wrong
Convert the accumulated feedback into a saved skill
Re-run the skill on future edits so it improves automatically
The loop demonstrated for turning Claude Code's first-draft ad edits into a steadily improving, reusable editing process.
Steal forany paid-social or content team producing repeat ad variations without a dedicated editor
CTA Breakdown
How they asked for the click.
VERBAL ASK
11:06subscribe
“drop a like comment what you guys want me to make more content about ... like, comment, sub, all that good stuff”
Soft, low-pressure ask at the very end after the value delivery, paired with a direct question to the audience about their own experience with the tool.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.
Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
A performance ad-agency founder breaks down the 12 static formats, messaging rules, and design levers behind ads that have spent hundreds of thousands of dollars, then shows the AI workflow he uses to research and draft them.
A creative strategist wires Claude Code to Slack, Notion, and an ad-research tool to build a self-updating brand brain that mines real reviews for ad copy and files its own reports.
A 38-minute walkthrough of a nine-step agency workflow that turns a curated swipe file, an ad-data MCP and an image model into hundreds of launch-ready static ads a week.
Paul J Lipsky hands his entire YouTube edit, silences to motion graphics, to Claude Opus 5.5 running through an MCP-connected editor, and walks through exactly what still needs a human pass.
A five-tool stack — transcription, cuts, AI b-roll, code-built motion graphics, and a self-review loop — that lets an agent finish a rough cut to a client-ready render without a human in the loop.
A professional motion designer hands his entire YouTube edit, script to graphics to quality check, to Claude Code and breaks down what it actually costs.
00:12
00:21
00:29
00:46
00:53
01:03
01:12
01:29
01:38
01:46
01:55
02:00
02:12
02:20
02:29
02:37
02:46
02:54
03:03
03:11
03:20
03:28
03:37
03:45
03:55
04:02
04:20
04:28
04:37
04:45
05:02
05:11
05:19
05:28
05:36
05:45
05:53
06:02
06:10
06:27
06:36
06:44
06:53
07:01
07:10
07:19
07:25
07:36
07:53
08:01
08:10
08:18
08:27
08:35
08:45
08:54
09:01
09:17
09:26
09:35
09:43
09:52
10:01
10:09
10:18
10:26
10:35
10:43
10:53
11:00
11:09
23:52
20:45
38:03
12:38
Video of the Day17:00
09:44