21:46Sirio · Tutorial
Dear Higgsfield, I Reverse-Engineered You in 12 Hours
A creator strips the 'who' out of a video to beat Seedance 2.5's face and music filters, then automates the whole workaround into a free 11-step agent workflow.
October 9thThe argument in one line.
AI editing gets reliable when the teleprompter records structured data about every take, so the cut comes from deterministic files instead of an AI guessing from raw footage.
Read if. Skip if.
READ IF YOU ARE…
- A creator who posts long-form video and wants to batch record and hand a near-finished cut to an editor or to an AI agent.
- A video editor worried about being replaced who wants to see how to run three times the output for the same pay.
- A Claude Code user looking for a concrete example of agents, skills, and deterministic files working together in one production pipeline.
SKIP IF…
- You want a click-by-click Premiere or DaVinci tutorial. This is an architecture walkthrough of a custom Python, Chrome and FFmpeg stack.
- You don't record talking-head video from a script, since the teleprompter data trick depends on reading scripted lines.
The full version, fast.
Sirio built a teleprompter that logs every take, retake and word timing, then feeds that data into a custom editing app that Claude Code drives. Because the cut comes from a deterministic edit plan, assembly errors drop to about 3% instead of the usual 20 to 30%. Seven steps run from recording to cutting with Whisper checks, planning boards, designing them from a 100-template system, animating to his voice, rendering with Chrome and FFmpeg, and writing the description. A 15-minute video costs about 40 cents in AI tokens and an hour and a half of human review, versus $750 to $900 and six to eight hours for a traditional edit. The editor's role shifts to creative director doing quality assurance.
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 →Lines worth screenshotting.
- AI clipping from raw footage and a transcript has a 20 to 30% error rate, while cutting from teleprompter-logged takes drops it to about 3%.
- A teleprompter already knows your script, your pace, every restart and the second each word is spoken, so that data only needs to be saved.
- The retake rule is deterministic and not AI: if a line is said several times, keep the last one.
- Saving instructions as JSON lets an agent read them exactly, with no guessing about what the cut should be.
- Without a teleprompter, you can still build the same deterministic system from your microphone audio.
- Boards are planned by asking one question per line: can the viewer understand this just by listening, or do they need a visual?
- Every motion-graphic board is an HTML file, so a library of 100 plus templates makes turning a new script into visuals easier.
- Chrome records each HTML board frame by frame and FFmpeg turns the frames into an MP4, with a 15-minute video taking about 20 minutes to render at 2K on an M3 Mac.
- Fifteen minutes of motion graphics cost about 40 cents of compute on a $200 per month Claude plan, against $750 to $900 from a human editor.
- Edit time for one video fell from five or six hours to about an hour and a half, and the editor finishes at least three videos in a normal shift.
- The editor's paycheck stays the same while output triples, so the gain shows up as three times the videos for the same money.
- Editors who charge per video can take three clients in the time they used to serve one, turning $700 a day into over $2,000.
- About 90% of the work is done before the human editor opens the project, leaving the last 10% as finding mistakes and directing the AI.
- The slowest part of the whole pipeline is now the creator sitting down to record.
- The editing app keeps all takes on a timeline matched word by word, so deleting a word also deletes it from the video.
Give the agent data, not footage to guess from.
WHAT TO LEARN
Capture structured data at the moment of recording, and an AI agent can cut, design and animate from a plan instead of guessing.
- Log every take, retake and word timing while recording, so the edit plan is a fixed file and not an AI's interpretation of the footage.
- Use simple deterministic rules wherever possible, such as keeping the last take of a repeated line, and save AI for the creative parts.
- Write agent instructions as JSON and markdown files in the project folder so any collaborator, human or AI, can pick up the work.
- Plan visuals by asking whether a line can be understood by listening alone, and design only the lines that need a picture.
- Build a reusable template library, because the more boards you have, the faster a new script becomes finished graphics.
- Keep a human in the loop for the last 10%, spotting boards that land early, cuts that feel off and sound that sounds flat.
- Measure the result as time, money and role, because the same system gives editors three times the output at the same pay.
Terms worth knowing.
- Take.json
- A per-take log file written by the teleprompter that records the lines read, restarts, retakes and the start and stop time of each attempt.
- Edit plan
- A deterministic file that says which take to use for every line of the script, keeping the last attempt when a line was repeated.
- Agents.md
- A note in the project folder that tells Claude what the files are and how to use them, for the AI or a human collaborator.
- Board
- A motion graphic overlay shown during the video, built as a small HTML page from a design system template.
- Board brief
- A markdown file produced by a planning skill that lists which lines need visuals and what type, used as instructions for the design step.
- Deterministic cut
- A cut generated from predefined data and fixed rules, so the same inputs always produce the same edit.
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.
This entire pipeline that you're seeing on screen is my AI video editing system. 90 % of it is handled by AI, and I built it from scratch for me and my editor so we can scale our content production, which means that I can start posting long -form content every single day. This is what I'm using from now on, and today I'm going to show you what is happening behind the system and why it changes three things for you, whether you are a creator or an editor.
It changes your time, your money, and your role, and I call these three shifts. Here's everything that's inside the free package for you to download today. For the teleprompter, it follows your voice, records your mic, and keeps track of every take and retake, every mistake.
An editing app that reviews and cuts your footage. You also get the agent instructions and the tools that transcribe your recordings, that compare your takes and check the results for you. Then finally, the design templates for claw design, the animations and the motion graphics so that you can build visuals around your video like I'm doing right now.
And yes, this video that you're watching right now is edited completely with this pipeline and the human input was only 10 % of the work. So let's put things in perspective for a second. 15 minutes of motion graphics is done in an hour and a half with this pipeline, which costs, believe it or not, about 40 cents in AI tokens.
Now, a normal human editor would charge me. anywhere from 750 to $900 for the same thing. And it would take anywhere from six to eight hours to deliver it.
Now, my system though works a bit differently from what you've seen online because it all starts with my teleprompter.
Hey friend, welcome back. If you're new here, my name is Sirio. I architect AI workflows and systems.
I deploy APIs and build AI businesses in the media space. And last year we made about $6 million doing just that. And just so we're in the same page, everything that I'm about to show you today is free on my GitHub.
You don't have to worry about how to recreate this prompt by prompt because you can copy everything right now. And the link is in the description. So instead, I'll just show you what I build.
and why. And I really want you to pay attention to the thinking behind it, because in a world where everyone can use AI to do pretty much anything, what really stands out is the way that you think and how you turn your knowledge into a product. So here's what we're going to cover.
First, I'll show you the whole system in seven steps. That includes the recording on my teleprompter, turning my teleprompter into a data collector, sending that data into an editing app that I built to sync with the teleprompter, and seeing how AI agents understand and my video project and cut it like a human would with only 3 % errors in assembling and clipping the long form video together.
Then we'll talk actual numbers. How long does it take? What does it cost?
What does it save? And how can the system scale beyond just me as a creator who uses social media to build AI businesses? What can we do with this?
Then we'll go through the three shifts that I mentioned in the beginning of this video one by one. So by the end, you will have my exact system. You can just copy it just like that.
I'm not asking literally for it. anything back in return so here's how the whole thing started okay i was sitting here and recording looking at my teleprompter like i'm doing right now and i was using an app that's called speakflow which basically turns your ipad into a prompter um and i was like wait a second like this thing already knows my script and it knows exactly which line i'm reading and how i'm reading it not not my words only but how fast i talk how how loud i am like every time that i mess up everything every time that i start a line over and the exact second that I would say each word and even when I go off script, right?
So this is data that's being collected, but is it actually being collected? This data is available to be collected. So I thought, okay, so why don't I just build my own teleprompter that can save all of that and send all the data that it logs straight into an editing app that I'm going to build too, synced to the prompter that's going to cut my video based on everything that knows so far about my script and how I speak.
So I wanted to test that out because in theory, like that would make so much sense. So I built both. With Claude, I build a teleprompter that saves everything while I record and an editing app that reads all of the data and cuts the video and adds all the motion graphics and syncs everything to my voice like you're seeing right now.
Whatever's happening right now is the product. Is that how I say that? It's the fruit of my labor.
Now the whole point of this is to scale my content production, right? So I can even give the AI editing app that I built to my editor so he will not have to like do all the repetitive stuff himself so we can move fast. So the moment that I stop recording, like if I do this right the second, my editing app knows exactly where to cut.
It's basically, this is the whole trick because it brings the error rate of like the clipping and assembly from like traditional ways of clipping your content through AI down to around 3%. That's normally 20 to 30%. if you just use AI to clip things and ask Codex or Claude to take a video and transcribe it and start clipping.
And here's how the whole thing happens in seven steps, okay? Step one, I record myself reading a script on my teleprompter and my Mac controls it. And my iPad shows me the words, so right here in front of the camera, and they sync in real time as I read, like you're seeing right now on screen.
And look at this, like what I can do is that I can change the text size. I did on my computer is going to change on my iPad. I can change the column width and I can even scroll myself and make sure that I can sync to whatever line that I'm reading.
And it syncs perfectly to my iPad. But normally I do not touch this because it listens through my mic and moves along with me as I talk. And while it does so, it's also recording my voice at the same time straight for my mic while my camera records the video on its own.
So camera is not really connected to the setup, but we are going to have to sync the audio later with. script under the hood this is very simple it's a little Python server that's running on my Mac and the teleprompter itself is just like a web page on Chrome and it follows my voice with a speech recognition that's already built into Chrome so there's nothing extra for us to install and the iPad just opens the same local host page which is how they sync because they go through the same Wi -Fi and the same router in the same server now when it's done recording it saves everything into one folder and this is really important because this folder has all the information that the editor will need later whether that's my human editor or that's like an external team member or just the editing app that i built so let's go into that folder and see what's inside for every take there's a file that's called take .json okay which is basically the diary of of our takes so there's all the data in here that we talked about and this is what it actually looks like if we open that say that i mess up this line and um i just start it again like let me do this live say that i mess up this
line and I'm gonna start this again limit let me do this live okay so now you can see on screen right here it catches that I repeated the the line twice and it marks it as retake all by itself I didn't press any buttons and this is saved in the take that JSON file and you can see both the tries and when each of them started and stopped but then there's one more file it's called edit plan JSON which basically is the plan for the cut it says which take to use for every line again it's talking it's communicating with our editor not the human editor in this case with our app with the editing app that we've built and this part is not ai because the rule is very simple it's deterministic if i say a line twice three times four times whatever that is it keeps the last one and then we have a file Obviously, that's called agents .md, which is just simply a note for Claude that explains what's in the folder and how to use it when you set it up.
But also if your human editor or like your friends are going to set up the same project, that's what's inside the folder. So now whoever has access to my editing app can use this information to automatically cut and assemble the footage. So let's talk about cutting the footage now.
Step two is where the cutting happens in the editing app that I build. And the editing app is basically a child of my teleprompter. because I literally used the teleprompter project to build on top of that.
So that's how the editing app came about. And here's what happens. Now, Claude.
It reads the agent's MD file first. If you remember, the agent MD file lives into the folder that the teleprompter produces. Then it syncs my audio from the prompter, which we recorded, to my camera footage.
Then it pulls up the edit plan, the edit JSON file, and follows it to grab the right take for every line. Then it starts the editing app on localhost. It all runs local in Chrome, and it uses FFmpeg, which is free, to actually cut and export the video at the very end.
And the reason that all of these things are saved as JSON like all the instructions is that the agent can read it exactly with no guessing so that means that the cut is actually coming from predetermined data instead of an AI watching like an hour of footage and transcribing it.
And then you hope that AI is going to make the right cut. We all know that that's not going to be the case. So the difference between what they teach you online and what I'm showing to you today is this.
And I think that that's a great hack. Even if you don't have a teleprompter, you can just use your mic and build the same the same deterministic system. Okay, once everything is put together, meaning that the agents have found the right cuts and they've lined them all up, it uses Whisper.
which is also free to transcribe the final cut that the agents put together. So he double checks that it actually makes sense and that every word can be heard. It's not like cut in the middle.
And then it finds all the pauses, all the ums, and removes everything that does not fit my instructions. And he does so by simply analyzing the audio waves and sends, as we said, the audio is synced to the camera footage. If you make a cut to the audio...
That cut is also going to happen to the video too. Everything is linked together. Now, a very important part, it does not just merge and cut everything into one file.
It puts all the takes on a timeline. So if something's off, the human editor can still fix it. Even if it's word by word, because as you can see right there, it matches the words to the footage.
And if you highlight and delete a word, it does the same thing in the timeline. So that's step number two. Step number three is the boards.
the most interesting part of all of this because it is saving me thousands of dollars. The boards are these graphics that you are seeing everywhere like behind me and I'll show them to you in a second. But first the boards are planned by a skill.
And the skill reads the whole final transcript from our cut previously from step number two, line by line, and figures out which lines need visuals and what type of visuals, right? What are we building here? And for every line that I speak, the skill asks one question, which for me in this case is, can you understand this just by listening or do you need?
And everything that the skill finds using all our agents gets saved as a simple markdown file like this one that you're watching right here, which is our board brief. And it's simply instructions for the agent to start designing. It's not the actual thing.
Step four is cloud code that's going to design every board for my own design system. which I spent a long time putting together. There's more than 100 boards here that have design styles, charts, graphs, prompts, and all the animations.
And the more boards or templates that you have, the easier it is to turn the new scripts that you are going to give Claude into graphics. Now, every board is just an HTML file. Basically, it's like a little web page.
It's just a very well -designed web UI. So how are the new boards actually designed? I give Claude the board brief from step number three and my own design.
System with all the templates so that every board follows a plan and they have the same design rules and nothing is left to chance Step number five all the boards now go through another skill that I build which tells Claude how things should Animate okay, for example, this one is one instruction on how it should use keyframes It only knows what you feed it right and then how you feed things So it goes through the skill and then animates everything and once that's done The animation gets synced to my voice with a script that happens in the back end You don't really see the process for example when you see that I point up Like this something shows up right when I say it right so that's the audio to graphics sync step number six We have to turn everything into a video so that we can export it and then publish it on on YouTube or any of your platforms Now remember that all the boards are still HTML.
They are code. So now what's gonna happen is that Chrome is gonna open like each animation, each board, and it's going to record it frame by frame, and FFmpeg is going to turn all of it into an MP4 video.
Now, on my Mac, a 15 -minute video takes about 20 minutes to render at 2K. I have a Mac at M3, and if your computer is actually slower, you can just rent GPUs on Rumpod. Then I watch the final video, and I approve it.
Step seven, a skill writes a description and the chapters of the YouTube video, in this case, using the transcript that I already... approved and that's it but how much does this really cost let's do the math my cloud plan is $200 $200 a month so that's about 27 cents an hour so this 15 minute video cost me about 40 cents for for the AI compute.
Now a normal human editor at $150 an hour for five to six hours, this is how long this will take, is about $750 to $900. And even if you pay your human editor the same exact rate, like $150, to run the system that I build or that you are going to build, that is about $225.
So that is 70 to 75 % less than what you would normally pay. You're saving around $500 to $700 on a video, which that's like $25 ,000 to $32 ,000 a year if you post. Four videos a month, let's say.
Now, if you're looking at this, you're either very excited or you're scared. And I know that a lot of editors think that AI just means that they are going to get replaced. And honestly, I get it.
You look at all these headlines. But all of them are clickbait or some of them are sponsored by the company that's selling a tool that's supposedly like replacing you. But I wanted to see what actually happens with my own human editor.
And that's where three shifts come from. I think it's his time, his money and his role. And I'm saying human editor just because I'm referring to an other like AI editor in this video.
It's very weird to refer to people as human editors, but just to make things extremely clear So the first thing that's changing is time if you're an editor an edit that used to take five to six hours now It takes about like an hour and a half So if you look at this graph, the gray bar is how long one video used to take Just five to six hours, and the blue bar is now.
And that dotted part is the four hours that he gets back on every single video. So now what I can do is that I can batch record a few videos in one day, and my editor can finish at least three of them in a normal shift. And funny enough, now the slowest part of this whole...
Right, the slowest part of this whole system is me sitting right here and recording. I'm sorry if I'm not extremely clear with what I say and how I pronounce things. English is not my first language.
But let's get back to the second shift, which is money. And this one is interesting because I still pay my editor, human editor, exactly the same as I did before. But he's putting out way more videos in way less time with way less effort.
If you look at this graph, the gray bar is before when... normal eight hour day my human editor would finish about one video and the blue bar is now same exact pay he finishes at least three videos so now this gap between the two bars that's the whole point that means three times the output For the exact same money he's paid the same but each video takes him a fraction of the effort So for the time that he's actually putting in he's earning way more I also get more videos for the same price and honestly we both win now if you're an editor That's a deal for you because you're gonna get paid per video or per month if you're gonna get paid by the hour That just means that a faster edit It's going to be a smaller paycheck for you.
And here's how you can actually scale this. Say that you charge seven hundred dollars for a 30 minute video or like 45 minute video before you could finish about one in one day. Right.
So it's going to take you like six to eight hours. Now you can finish three. So instead of one creator, you can take on three in the same exact hour and you start automating the process.
And that's seven hundred dollars that a day that you were turning into over like two thousand dollars without having to work a single extra hour just because you can handle more output and and let's say that you're full now you can hand the same exact system the same exact workflow to maybe a junior editor and teach them what good looks like and the third shift which i think that matters the most here is your role before my my editor made everything by hand right cutting every take building every board keyframing everything now about 90 of that is done before he even opens a project um and the cut cut that's done by the pipeline, comes in and around with like 3 % errors.
So his job, like the last 10%, is finding the mistakes, fixing them, and directing AI to do a better job. And that's just what takes no more than an hour and a half. And the great part is that you don't really have to be technical.
You don't really have to learn how the AI works. I mean, obviously, if you do so, that's way better because you're going to optimize the system. But I figured the system out for you.
I figured the system out for him in this case. So if you are editing, instead of executing the edits yourself, you're becoming more of like a creative director doing quality assurance. Like your job is spotting the boards if they land too early, if the cuts feels off, if the sound sounds flat, all these minor little details that are extremely important for creating quality video.
And now this is what I'm seeing in my own business. It's not a study, but this is very clear to me. And take this from someone who knows the space.
Guys, I actually want to retire like tomorrow. I'm very tired. It's nuts in here.
Okay, I know that this was a lot. So if you made it all the way to here, thank you. Because here's what I really wanted to take from this.
I had a problem. My editing was way too slow. It was too expensive.
And I just built my own solution to it. I did not wait for an app to add a feature or someone to like release an MCP because everything that I need was already available right here, right? Actually right in front of me.
And what made the difference was how I architected this entire thing and how well I understand the tools and how I put them together into a system that works for me. It has to work for me. And if you do not need the system yourself, you can do it for other people.
You can package it up. You can sell it to agencies, to other creators or students. And to run this, literally, you just need a camera.
a mic, a tip like an iPad for teleprompter, like you can find alternative solutions to that Chrome cloud code. Literally, that's it. Now, the whole thing is free on GitHub.
The link is below. And to install it, just like copy the link, paste it into your cloud code and tell it this read the start here guide in this repo, install everything it lists and run the setup check. So my friend, take the system running on your next video, change whatever doesn't fit for you and see where that takes you.
Because being curious, that's enough to make it out there. Now, if you want to build systems like this with other people, get feedback, just like hang out with me, join Public AI. It's my free community.
The community is for people who want to go further than just making AI content. It's actually for people who want to build something like a media company, an app, agents, whatever pushes the boundaries of what you think is possible. So my friend, a piece of advice, do not just watch tutorials on how to use someone else's software.
Start building your own solution. The link is right below. Join me today and do not forget create without limits.
This is Syria.
The Hook
The bait, then the rug-pull.
Most AI editing demos ask a model to watch footage and guess the cuts. Sirio flips it: the teleprompter logs every take as data, so the agent cuts from a plan and not from a hunch.
