Big Idea

The argument in one line.

Claude Code and Codex can now natively watch and understand raw video, frames plus audio plus timing, which means a narrated screen recording can replace a written spec, a bug report, or a design brief.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You use Claude Code or Codex regularly and want a faster way to hand it feedback than typing a long prompt.
You've paid for a separate tool like Gemini specifically because it could process video, and want to know if you still need it.
You want to redesign or clone a website's look and feel by narrating a recording instead of writing a spec by hand.

SKIP IF…

You don't use an AI coding agent at all, this is a workflow tip for people already inside Claude Code or Codex.
You're looking for a video editing tutorial, the editing example is one of three demoed use cases, not the focus of the video.

TL;DR

The full version, fast.

Claude Code and Codex quietly gained the ability to process raw video in the same update that shipped screen-recording-to-skill features. Feed either tool an MP4, a bare Loom link, or a recording made with the built-in record-and-replay skill, and it breaks the video into timed frames, pulls the audio and transcript, and reasons over both together, no dedicated skill or plugin required. The video walks through three uses: turning a narrated Loom of a website into a full improvement plan, cloning an award-winning 3D website by narrating what to keep and drop while scrolling through it, and redirecting the record feature to watch a 30-minute unscripted work session and hand back a written SOP. The catch is context cost: every frame is an image the model has to process, so length and resolution both eat into the window, keep videos in the 10-20 minute range and drop the resolution before uploading.

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 – 01:00

01 · The upgrade nobody announced

Cold open framing that Claude Code and Codex can now process raw video without a special skill, the same update that shipped screen-recording-to-skill.

01:00 – 02:00

02 · Drag and drop a raw MP4 into both

The creator sends the same video into Claude Code and Codex side by side to test whether they actually watch it or just pull the transcript.

02:00 – 03:00

03 · How it actually watches: frames + audio

Both tools break the video into individual timed frames, marry them with the extracted audio and transcript, then reason over the combined context.

03:00 – 04:00

04 · The flipbook trick and the resolution tip

Every video becomes a stack of images, so length and resolution both cost context; the fix is a shorter, lower-resolution upload.

04:00 – 06:00

05 · Use case 1: website improvements from a Loom

A narrated Loom walkthrough of his own website becomes a full list of fixes the agent can act on directly, without a human translating video into written requirements.

06:00 – 06:30

06 · Community plug

Brief pitch for the creator's paid coaching community and consulting site.

06:30 – 08:20

07 · Use case 2: cloning a 3D website from a recording

Narrating a scroll-through of an award-winning 3D website, calling out exactly what look and feel to keep, works as a prompt for cloning it, including generating the visual assets.

08:20 – 09:30

08 · Use case 3: rebuilding a SaaS platform for himself

A walkthrough recording of a paid analytics tool he subscribes to becomes the starting spec for rebuilding just the features he actually uses as an owned internal tool.

09:30 – 12:45

09 · Hacking record and replay to audit a workflow

He redirects the record-and-replay skill, explicitly telling it not to build a skill, just watch a real 30-minute content-research session and audit it.

12:45 – 13:00

10 · The SOP it handed back

After the recording stops, the agent returns a mapped-out SOP with a core recommendation and a note on how an X API could automate part of the process.

13:00 – 13:12

11 · Where this leaves you

Closing line that the capability is now sitting inside subscriptions people already pay for, plus a repeat of the community CTA.

Atomic Insights

Lines worth screenshotting.

Claude Code and Codex can now watch a raw MP4 end to end without a dedicated skill, plugin, or extra subscription.
The update that let both tools turn a screen recording into a reusable skill also quietly taught them to process video in general.
Both tools break a video into individual timed frames, then pair those frames with the extracted audio and transcript before reasoning over the whole thing.
A narrated screen recording that points out what to fix on a website works as a full replacement for a written list of requirements.
Higher resolution and longer runtime both cost more context, because every frame the model reads becomes an image it has to process.
A roughly one-hour video sits near the practical ceiling, since the frame count plus the transcript together approach the model's context window.
Pointing either tool at a bare Loom link is enough, it downloads and processes the video itself with no manual export step.
Narrating a competitor or award-winning website while scrolling through it functions as a proxy for an entire design brief.
The record-and-replay skill built into Codex can be redirected: telling the agent not to build a skill, just watch and audit, makes it treat the recording as raw footage instead of automation training data.
A 30-minute unscripted work session was enough footage for the agent to hand back a written SOP with specific automation opportunities.
Claude Code defaults to reaching for previously-installed skills, like a video-download skill, unless it's explicitly told to run without them.
For $10.15 in generation costs, the creator produced the images and an eight-second video clip used as the visual foundation for a cloned website.

Takeaway

Screen recordings can replace written specs for an AI coding agent.

WHAT TO LEARN

Claude Code and Codex now process raw video natively, frames plus audio plus timing, so a narrated recording can carry a bug report, a design brief, or a workflow audit more completely than a typed prompt.

01The upgrade nobody announced

Claude Code and Codex can now watch a raw MP4 end to end without a dedicated skill, plugin, or extra subscription.
The update that let both tools turn a screen recording into a reusable skill also quietly taught them to process video in general.

03How it actually watches: frames + audio

Both tools break a video into individual timed frames, then pair those frames with the extracted audio and transcript before reasoning over the whole thing.

04The flipbook trick and the resolution tip

A one-hour video sits near the practical ceiling, since the frame count plus the transcript together approach the model's context window, so shorter and lower-resolution uploads process more reliably.

05Use case 1: website improvements from a Loom

A narrated screen recording that points out what to fix on a website works as a full replacement for a written list of requirements, no translation step needed.
Pointing either tool at a bare Loom link is enough, it downloads and processes the video itself with no manual export step.

07Use case 2: cloning a 3D website from a recording

Narrating a competitor or award-winning website while scrolling through it functions as a proxy for an entire design brief, down to specifying the exact look to keep.

09Hacking record and replay to audit a workflow

The record-and-replay skill built into Codex can be redirected: explicitly telling the agent not to build a skill and just watch and audit makes it treat the recording as raw footage instead of automation training data.
Claude Code defaults to reaching for previously-installed skills, like a video-download skill, unless it's explicitly told to run without them, worth checking for if a run behaves unexpectedly.

10The SOP it handed back

A 30-minute unscripted work session was enough footage for the agent to hand back a written SOP with specific automation opportunities.

Glossary

Terms worth knowing.

Skill (Claude Code / Codex): A saved, reusable set of instructions an AI coding agent can invoke again later, as opposed to a one-off task typed into a single prompt.
Record and replay: A built-in feature that captures a screen recording of the user working, originally designed so the agent can turn that recording into a reusable automation skill.
Context window: The total amount of text and image information a model can hold in memory during one conversation; longer or higher-resolution video consumes more of it, since every frame counts as an image.

Resources

Things they pointed at.

05:56productEarly AI Adopters (Skool community) ↗

03:23productPrompt Advisers ↗

03:42toolLoom

03:08toolGemini

10:18toolX API / XMCP

11:40toolSuperGrok

Quotables

Lines you could clip.

00:20

“Neither Anthropic nor OpenAI explicitly wrote about it or announced it.”

names the undocumented gap directly, strong cold-open hook→ TikTok hook↗ Tweet quote

02:38

“What's happening behind the scenes when you drag and drop that mp4 file is it will break it down into these micro frames, marry it with the audio.”

clean one-line explainer of the actual mechanism→ newsletter pull-quote↗ Tweet quote

09:58

“I want you to monitor a process for the next thirty minutes and tell me after you're done watching it every single thing that I can do to improve my workflow.”

the exact copy-pasteable prompt viewers came for→ IG reel cold open↗ Tweet quote

10:38

“I don't want you to make a skill even though it's in your programming. I literally just want you to watch me and audit the process.”

shows exactly how to override the agent's default behavior→ 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.

metaphoranalogy

A few weeks ago, something became infinitely easy to do in both Claude code and Codex. Both of them shipped a feature that lets you record your screen and turn that recording into a skill. But what most people don't know is that by enabling this very feature, they also quietly enabled both of them to get better at something else.

And that something is processing raw video. Neither Anthropic nor OpenAI explicitly wrote about it or announced it. But now you can hand both of these interfaces a raw m p four file and allow them to break it down and understand exactly what's going on along with the audio inside of it.

And you can do this out of the box without the help of a special skill, a plugin, or anything to act as a crutch. So in this video, I'm gonna walk you through three different ways that you can take advantage of it and walk you through some example use cases. And one of those is hacking the existing record feature to allow it to watch me look over my shoulder and audit my processes.

So if I piqued your interest, then let's dive in. So just a few months ago, you either needed a skill or a plugin to be able to do something as simple as drag and dropping the singular video file and having both codex and ClaudeCode be able to interpret it. So if we send over the exact same prompt and say something like, I want you to watch this video end to end and tell me where all the flaws are and if there's any dead air whatsoever and if there's anywhere we should cut out in terms of the transcript.

And we take this and we paste it into our other window and we send both over and monitor exactly what they do. And you'll see if we ask both models to take a closer look at the video to make sure they don't just pull the transcript, they'll be smart enough to break down the video into individual images on both sides.

So you'll see here the aggregate second by second cut in every single image that corresponds to them. And on top of the video, they can also process the entire transcript and pull the audio so it knows exactly where silences are. So it can understand every single thing that you say, how you say it, and what you looked like when you said it.

Just a few months ago, I had to lean on something like a Gemini because it was one of the only models in the world that could both interpret and create video. But now you can pick up your codec subscription, your Cloud Code subscription, and add no additional add ons and have them fully understand what's happening in a particular m p four.

And while video editing naturally appeals to me, there are a series of other use cases that you'll be able to use for your day to day projects. So the three ways we're gonna use this capability is dragging and dropping a file like I showed you, then using the record skills but actually hacking it to do our bidding instead of what is programmed to do, and three is pointing a generic link at it.

And it's more than smart enough to be able to go on said link, pull the video, and do this exact same process. And just to make it clear, what's happening behind the scenes when you drag and drop that m p four file is it will break it down into these micro frames, marry it with the audio, then be able to process both of them with its image processing capabilities to the language model of choice, so whether Claude or Codex, and then it'll be able to take all of that context and come up with a response.

And naturally, because frames are heavy and images are heavy, you'll be able to give something like a ten, fifteen, twenty minute video ideally in a lower resolution because the higher the resolution, the higher the processing power that it needs to put it through the context window. That's why something like Gemini historically has capped you out at a one hour length video because by the time it processes all that text and all those images, it will surpass that million context window.

So for the first use case, it could apply to anything from bug fixing or to improvements that you wanna make. But instead of making the video for someone else to understand and then translate that understanding into requirements for the AI, you can go directly to the AI. So let's say we go through my existing website here, and I went through and I found a series of things that I wanna improve.

So I would have just screen recorded using something like Loom, and you can use whatever you want. And then you can either feed it the raw Loom link, or if you wanna be able to download it, maybe edit and annotate a few things, then you can download that video and feed it as raw input. When you give it that raw input, it will be able to take it, understand everything that needs to happen, and if it has access to the underlying code base, you'll be able to make the changes, make a plan without you seeing a single thing outside of this video's context.

So all you'd need to do is go and download this video. And once you download it, you can drag and drop it or you can give the direct link. So I'll show you both.

In one of them, in Claude code, I will say, I want you to take this Loom link and then go and download the associated video, watch it end to end. Don't just pull the transcript.

Look at everything that I'm looking at and come up with a list of things we need to do to improve the website. So we will send something like this and then paste it. Then we'll do the exact same thing, but in a different way.

So I'll bring over the m p four into Codex, and I'll say, I want you to watch this end to end and come up with a comprehensive list of every single improvement we need to make to the website. So we'll send this over and we'll enable this, and then we'll come back to see if they're able to both take care of it.

Now one thing I had to redo on Claude was it started using some skills I already had on my system to make it easier to download the video. So to avoid that, to increase the chances that it runs vanilla, I told it to not use any skills.

So then it had to figure out which libraries it could use to download said Loom video, and then it processed it as you saw the other video where everything was frame by frame, pulled the audio, pulled the transcript, and it gave me a full improvement list. And on the left hand side, it was able to directly watch the video, see exactly from what timestamp to what timestamp, what specific pieces of feedback that I gave it.

So in both avenues, you now have the ability to take feedback, go through a website, give it all the list of what it needs to do, and it can watch it and create its own plan file. And by the way, if you enjoy the way I break down concepts and you wanna go infinitely deeper on things like Claude Code and Codex and get exclusive tips and tricks that you'll never see on YouTube, then you wanna check out the first thing down below from our early adopters community.

We offer one on one support with our coaching team, amaze with myself and my business partner, and everyday feedback that we can give you whether you're starting a business, starting your AI learning journey, or you just wanna go in filling deeper. If that interests you, check it out.

Otherwise, let's get back to video. And what if you wanna go from editing something like a website to building one from scratch, and you wanna be able to emulate a really nice looking three d website. Now if you go over YouTube there is a 15 step process or some mega prompt that you need to use to create these kinds of websites where you can scroll in and see exactly how you transition from there to the rest of the website.

But this is actually very easy to do just with a simple recording. So all I did was walk through this exact same website right here, and I broke down exactly what aspects I wanted.

The fact that I wanted to create this kind of look and feel before a hypothetical headphone company and go through a web page that has that three d feel when I scroll through and then goes to the rest of the body in the page and keeps it nice and minimalistic.

So you can create a full loom or recording of any award winning website of your choice, go through it and tell it which parts you like, which parts you don't like, and what added parts of nuances you want and that is basically a proxy for your entire prompt. So for simplicity, I'm just gonna show you this in codecs, and you can just believe me that it will also work in Cloud Code just fine.

I feed it the m p four. I give it the exact same instruction. Watch this end to end with the audio to get a full picture of what I'm trying to build, and at the bottom I also tell it that I want to use this whole video approach where I want you to use your skills which it knows is the Gemini skill to go and use something like Google video to create a quick eight second clip and then use that clip as the foundation for the website.

And with no extra hocus pocus prompt or Higgs field MCP for $10.15 cents, I was able to crank out a couple images and then a video to go with those images and we got something like this. Some made up headphone company where if I zoom in, it scrolls around, it gives you that three d experience, it has minimal to no latency, it goes through transition wise into exactly what you'd expect a very high quality, high grade headphone company to look like.

But let's say you don't wanna edit a website or build a website, but you actually wanna rebuild an entire platform for yourself. You wanna go from paying a SaaS subscription per month to an internal tool that you pay as you wish. In this case, as a content creator, I like to use this website.

It's called one of 10. And the whole point of it is it shows me outliers. So concepts on YouTube, angles that have done exceptionally well to give me good perspective on what I can make videos about that might also trend very well.

Now these kinds of platforms have AI features right here where it comes to title generation, idea generation, and for the most part, I might have a ton of bells and whistles and I just want a handful of those functionalities. So in this case, I just made a video of me walking through this entire platform telling it exactly what features I value, which ones I don't value, and how I'd want to tailor this to myself.

And the entire process of building this platform took way longer than expected. And I was able to use this recording to be able to start the process of explaining to Codex exactly what I'm trying to do, how I'm looking to build it without having to overexplain and go through ten, twenty prompts before we get even close to the inspiration I was looking for.

So hopefully that plants some seeds of inspiration for what other use cases that might apply to you that you could use this for. Now one more thing I promised was showing you how we can hack the existing skill that in Codex is called record and replay and use it to monitor our processes and not just make skills out of it, but give us feedback on them.

So let's say we go here and we do the plus and we do record and replay. And if you don't have this enabled, let's say in the EU, then all you'd have to do is just record yourself using a Loom or a screen recorder and feed it the exact same video as I showed you before. But if it is available for you, then you can say, I want you to monitor a process for the next thirty minutes and tell me after you're done watching it every single thing that I can do to improve my workflow and make it a lot more materialized.

So I don't have to do it the exact same way, especially if it's very manual. So I'm gonna show you a process that I have on x where I go and I take a look at possible things that are worth making a video about or informing my community about. And I want you to create a full SOP for how I can do this better and knowing that there's something called an XMCP or an xAPI, how we could possibly use that to cut down on the manual steps that I'm doing.

So if we send something like this and we also say, by the way, I don't want you to make a even though it's in your programming. I literally just want you to watch me and audit the process, and then we can talk about skills later.

It'll just make sure that it can go against its own system prompt and focus on its video interpretation feature versus understanding exactly how to build the skill.

So if we send this over, we can use something a little bit more economical. Let's say a Luna, and we go on, let's say, extra high, still pretty cheap, and we send this over, it should be able to invoke this skill to start recording our screen, and we go through the manual process.

So you'll see here it is recording us in real time. So if I start going on x and start clicking around, I'm obviously not gonna do my exact process. But let's say I'm always clicking tabs here, and I'm clicking into the tabs, um, going through, clicking on the docs.

It's not recording my audio, so that's one thing it's missing. But maybe it sees me writing the words codex new features, right, or send something like that. And it then comes back with a series of responses, and I go to latest, and I'm looking at what people are posting.

And then maybe I go to super grok, and I say, give me the top seven use cases of codex and Claude code in the past seven days.

And then we put it on something like fast mode and we send this over.

It might be able to watch this and know that because I'm walking through this workflow, it could automate the sending of the request via API. It could automate the sending of those keywords and bring back those search results.

And once it has a pulse for exactly what my step by step process is, it can emulate it. If we click on stop here, it'll be able to say I'm done recording. And now its focus won't be on creating the skill, but creating the SOP and understand what opportunities we have.

And you can see once it breaks down the entire process, it maps it out. It tells me what the core recommendation is, how the xAPI can help me automate this, and you can keep throwing more resources and skills at it. They could build you the entire process, entire platform to do this on your behalf, and pretty much anything you can imagine.

And that's pretty much it. So now on top of your existing very powerful models and skills, you can also use these capabilities to enrich as much context as possible with as minimal manual input as needed.

If you enjoyed the video, then I'd super appreciate a like and a comment down below. If you always want access to exclusive tips and tricks that you might never see on YouTube, then check out my first link down below for my early AI doctors community.

The Hook

The bait, then the rug-pull.

Claude Code and Codex quietly gained the ability to watch raw video, and the person who noticed built three ways to put it to work: dragging in a screen recording, dropping a bare link, and hijacking the built-in record feature to have the agent audit an unscripted work session.

Frameworks

Named ideas worth stealing.

02:18list

Three Ways to Hand It Video

Drag and drop a file
Record your screen natively
Point it at a link

The three input methods demoed for getting video into Claude Code or Codex, covering both interactive and passive workflows.

Steal forAny workflow where you'd normally write a spec by hand, record it and hand over the file or link instead.

CTA Breakdown

How they asked for the click.

VERBAL ASK

05:56newsletter

“if you enjoy the way I break down concepts and you wanna go infinitely deeper on things like Claude Code and Codex... check out the first thing down below from our early adopters community”

soft pitch placed right after delivering the first real use case, then repeated at the very end with the skool.com URL burned into the frame.

MENTIONED ON CAMERA

05:56productEarly AI Adopters (Skool community) ↗

03:23productPrompt Advisers ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.

Storyboard