11:32Andy Diep · Tutorial
3 Tips for AI Editing in Claude Code + DaVinci Resolve
An unscripted walkthrough of the editing philosophy, and the actual Claude Code prompts, behind one creator's AI-assisted cutting workflow.
September 2ndThe argument in one line.
A rule file reverse-engineered from one creator's 900 published clips can auto-select and cut long-form podcast content into standalone Q&A segments inside DaVinci Resolve, but the same rules degrade on content they weren't built for, and a tool reporting success is never proof a human can see the result.
Read if. Skip if.
READ IF YOU ARE…
- A video editor or content-ops person who cuts long-form interviews or podcasts into short clips and wants to see a working AI-assisted pipeline end to end.
- Someone building their own Claude Code automation who wants a real example of writing an explicit rule file instead of one giant prompt.
- A creator sitting on hours of unedited podcast or livestream footage who wants a sense of what an AI clip-mining system can and can't do today.
SKIP IF…
- You're looking for a click-by-click tutorial — this is a systems walkthrough, not a step-by-step build guide.
- You need clip selection to work out of the box on a video format the system wasn't tuned for.
The full version, fast.
A ten-year editor built a Claude Code system that mines long podcasts for complete question-and-answer segments and writes the cuts directly into DaVinci Resolve. The rules came from reverse-engineering 900 uploads on Alex Hormozi's MoreMozi channel: a clip has to be one bounded question with a closing verdict, roughly 2-10 minutes, with timestamps snapped to exact transcript word matches rather than model-guessed times. Pointed at three unrelated podcasts totaling 7 hours 37 minutes, it produced 107 accepted clips in about 52 minutes of machine time, but rejected podcast dialogue as 'not a caller question' and threw out 41 clips under its two-minute floor since podcast exchanges move faster than the livestream calls the rules were built on. An earlier version also silently failed: DaVinci reported every clip placed successfully while all 12 timelines sat empty, because Resolve's default timecode starts at 01:00:00:00 instead of zero. Nothing publishes until he's watched every clip himself.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:15
01 · The payoff up front
Andy shows the finished clip wall, dozens of ready clips with titles and an 'Open in Resolve' button, and states the headline number before explaining anything: Claude cut over 100 clips from 7 hours 37 minutes of podcast footage in about 52 minutes of machine time.
00:15 – 02:32
02 · Kicking off the run
He shows the three podcasts open in browser tabs, copies a link into Claude Code, tells it to use the FAQ engine on the podcast, and starts a timer. He jump-cuts to the finished run, then opens the review board and clicks 'Open in Resolve' to show a clip land in a real DaVinci Resolve timeline.
02:32 – 04:42
03 · Reverse-engineering MoreMozi
He explains the system's rules came from scraping 900 uploads off Alex Hormozi's second channel, MoreMozi, and working backward from what actually got published, since he never had access to the raw footage those clips were cut from.
04:42 – 07:22
04 · Watching the system work
More demo footage: the review board showing title, thumbnail, and verdict reasoning per clip, and DaVinci Resolve timelines being built automatically from a Chris Williamson podcast interview.
07:22 – 09:47
05 · Why he built it and the scope of the test
He states his goal, helping editors bring AI into their workflow rather than replacing themselves, and lays out the actual test: three podcasts, Dave Ramsey, Diary of a CEO, and Chris Williamson, totaling 7 hours 37 minutes, with the four-hour Williamson episode as the bulk of it.
09:47 – 12:12
06 · How the selector decides
Hormozi's channel posts 20 to 40 times a day; only the long-form Q&A lane, 311 of 900 uploads, was rebuilt. He shows the actual policy.json rule file: one caller question equals one video, explicit length bands, and a requirement that a clip be one bounded question with a closing verdict.
12:12 – 14:38
07 · Proof of work: the scorecard verdict
He shows the FAQ Engine's scorecard listing accepted clips, then reads the literal reasoning the selector wrote for one: a 2:27 clip inside the 120-600 second band, one bounded question about founders misusing AI with a closing verdict. He also explains the mechanical cutting pass: measuring silence, keeping five-frame handles, and writing the range into Resolve.
14:38 – 17:03
08 · Where it wins, where it breaks on podcasts
He lists where the agent clearly wins, matching words to time, tightening edges, cutting silence, repeating the same operation across a hundred clips without fatigue, then explains the rule file was tuned to Hormozi's livestream caller format, so on conversational podcasts it rejected host-led asides and threw out 41 clips under its two-minute floor.
17:03 – 19:21
09 · The DaVinci timecode bug and closing lessons
He reconstructs an earlier failure where Resolve reported every clip placement as successful while all 12 timelines sat empty, caused by Resolve's default timecode starting at 01:00:00:00 instead of zero. He closes on where an editor still wins, packaging and titling, his rule that nothing ships unwatched, and points to the GitHub repo, a free consult call, and two linked follow-up videos.
Lines worth screenshotting.
- A Claude Code system trained on 900 uploads from one channel cut 107 accepted clips out of 7 hours 37 minutes of podcast footage in about 52 minutes of machine time.
- 99.4% of the clip timestamps it proposed landed on a real moment in the recording, not a timestamp the model invented.
- The system never trusts an AI-guessed timestamp: it proposes the first words of a question and the last words of an answer, then a plain text search finds those exact words in the timed transcript and snaps the cut to the real location.
- The rule file was reverse-engineered from 900 published uploads on Alex Hormozi's MoreMozi channel, because the creator never had access to the raw unedited footage those clips came from.
- Out of 900 uploads, only the long-form question-and-answer lane, 311 of 900, got rebuilt; the rest of the channel's format wasn't attempted.
- A publishable clip has to contain one bounded question with a closing verdict, land inside a 120-600 second window, and can't be a quote yanked out of context.
- Pointed at three podcasts the rules weren't built for, the system rejected host-led asides as 'not an independent caller question' because the source channel's format was call-in Q&A, not two people in conversation.
- 41 proposed clips were thrown out for falling under the system's two-minute floor, because conversational podcast exchanges move faster than the livestream caller answers the rule file was tuned on.
- A single stuck-together word pair in one transcript caused the system to lose almost nine minutes of a good answer it couldn't find until the transcription was rerun.
- DaVinci Resolve reported every clip placement as successful while all 12 timelines sat completely empty, because Resolve's default timeline timecode starts at 01:00:00:00 instead of 00:00:00:00, writing every clip an hour before the visible start.
- A tool telling you it succeeded is not proof that a human can see the edit.
- The remaining edge for a human editor is packaging: a system can surface a moment and write a title describing the topic, but knowing why anyone would click on it is a taste judgment only real viewers settle.
- None of the 107 accepted clips ship without the editor personally watching every one first.
Write the rule down before trusting AI with it
SYSTEM DESIGN
An AI clip-selection system only works as well as the explicit, written rules it checks against, and those rules quietly fail the moment the source content stops matching what they were built for.
02Kicking off the run
- A working AI editing pipeline can start from a single instruction: paste a podcast link into Claude Code and tell it to apply your rule file, then let it run unattended for tens of minutes.
- Running the same job across three separate long videos at once is a realistic way to test whether a system generalizes, not just whether it works one time.
03Reverse-engineering MoreMozi
- When you can't see how someone actually edits, you can still reverse-engineer their standard by scraping everything they've published and working backward from the finished pattern.
- Scoping the rebuild to one specific content lane instead of a creator's entire output keeps the resulting rules coherent instead of vague.
04Watching the system work
- A review interface that shows a title, thumbnail, and the model's own reasoning for each proposed clip lets a human approve dozens of cuts far faster than opening a timeline for each one.
- Writing accepted clips as their own timelines inside the real editing software, instead of exporting finished files, keeps every cut adjustable after the fact.
05Why he built it and the scope of the test
- Testing an automation against unfamiliar source material, not just the content it was trained on, is what actually reveals whether the underlying system is sound.
- The goal of AI-assisted editing is enhancing how an editor works, not replacing the editor's judgment on what ships.
06How the selector decides
- A publishable clip needs an explicit, falsifiable definition written down before you ask an AI to select anything, or you end up moving the goalposts after the fact.
- Duration alone is easy to check, but completeness needs its own test: a real question has an edge, and so does a real answer.
07Proof of work: the scorecard verdict
- Never trust an AI-proposed timestamp directly; separate the proposal from a literal text search that confirms where those exact words actually occur in the transcript.
- Requiring the system to write out its verdict reasoning in plain language, not just a pass or fail flag, makes each accepted clip auditable after the fact.
- Detecting silence and keeping a small handle of extra frames on each cut is a mechanical rule a system can apply identically across a hundred clips without fatigue.
08Where it wins, where it breaks on podcasts
- An AI system reliably wins at repeatable mechanical work: matching words to timestamps, tightening edges, cutting silence, and doing the same operation on clip 100 as cleanly as clip 1.
- A rule file trained on one format doesn't transfer cleanly to an adjacent one; it will reject valid content for not matching the shape it learned.
- Resist the urge to keep tuning the rules until they fit whatever you're testing on, since that just fits three videos instead of showing what the system actually does.
09The DaVinci timecode bug and closing lessons
- A tool reporting success is not proof the output is correct; verify by reading the actual result back, not by trusting the success message.
- The editor's remaining edge is packaging and titling, a taste judgment about why someone would click, which only real viewer behavior can actually confirm.
- Automate only the steps you can fully describe out loud as a repeatable rule; anything you can't describe that precisely still needs a human decision.
Terms worth knowing.
- FAQ Engine
- The Claude Code system in this video that scans a long podcast transcript, identifies complete question-and-answer segments, and writes accepted clips as timelines directly into DaVinci Resolve.
- Selector
- The part of the pipeline where Claude Code makes the actual editorial judgment: given a candidate question-and-answer chunk and the written policy, it decides whether the segment is complete enough to publish and writes down its reasoning.
- Timestamp snapping
- Instead of trusting an AI-guessed cut point, the system proposes the exact starting and ending words of a segment, then searches the timed transcript for those literal words to find the real timestamp.
- Length band
- The 120-600 second (2-10 minute) duration window a candidate clip must fall inside to be considered publishable under this channel's rules.
- Handles
- A small buffer of extra frames, five in this system, kept on either side of a cut so the edit doesn't start or end mid-word.
Things they pointed at.
02:32channelMoreMozi (Alex Hormozi's second channel)
Lines you could clip.
17:55
“A tool telling you it succeeded is not proof that a human can see the edit.”
18:41
“If you can describe the step out loud, you can probably stop doing it yourself.”
18:28
“Not one of those 107 goes anywhere until I've watched it myself.”
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.
metaphor
What if I told you Claude cut 10 of these clips, make that over a hundred without me doing anything in about 52 minutes across three videos, that's seven hours and 37 minutes of footage of podcasts. That's the power of using AI automations to help you edit. Okay.
So what I have on screen are a few different podcasts off of YouTube, right? This one is three hours and 14 minutes long. I'm going to take this podcast by copying the link.
We're going to go into Claude. I want you to use the FAQ engine to cut up this YouTube podcast. And you're going to see how long it takes in order to cut a three hour video podcast.
Start the timer both at the same time. All right, ready, set, and go. And .
All right. So it seems like we have finished everything. I think it's been done for quite a while.
I'll check back on the video itself to see when it finished. And so after it finished cutting all the clips, it's going to show us on an HTML page, kind of like a web page with every single clips with. title, a thumbnail, possible thumbnail, and even a button that we can give feedback on, or we can press a button and it goes into DaVinci Resolve so that we can fix it ourselves, whichever you like.
You see how many clips it did in just a short span of a time. By the way, this is a brand new feature. I have not tried this feature yet.
Oh. The point of this video is to show you that whenever you learn how to automate, you learn how to build the systems behind something that does this, like cutting videos. And here is our videos.
Let's listen to it. If you're an editor or a creator on the edge about letting AI touch your edit, I don't blame you. I've been getting comments like these.
On a 10 -minute talking head video, these comments are right. Yes, you can do that by hand. But on four hours of two people talking, it's a different job.
And at some point, just do it yourself. It stops being reasonable advice. This might be easy once it's all built out, but it wasn't as simple as just saying, hey, edit my videos.
This just isn't how it works. There's a channel on YouTube called More Mosey, which is Alex Ramosi's second channel. He has over 3 million subscribers.
And I really think it's... The kind of beginning phase of what editing with AI is starting to look like. It's not, hey, edit my full video and then you get back a final output.
It's something you have to refine over time. You have to study how it works based off of what other people do. So I scraped together 900 of their uploads and worked backwards from what actually got published.
Because I never got to see how anyone over there was cutting anything. All I ever had was the finished products. So what I built is a system that takes any long -form content, 10 minutes, 1 hour, 4 hour, and cuts it into digestible pieces.
Cloud Code, it does the judging, and DaVinci Resolve is where the cuts land. This video, it goes really deep. The mechanisms behind it, how I went about building it, and what it took to set up, how long it actually takes to run, and where it fell short on those podcasts.
And at the very end, I'll hand you the scripts so you can go build this yourself. Now, I've been doing this about 10 years.
My goal here is to help editors bring AI into how they work. Enhance how they work instead of replacing themselves with it.
It needs to happen. Find a way to make this happen. And if everybody else gets up to that standard, they're just not built for this particular company.
They're built for someone else that makes a mediocre product, but they're not built for you. Okay. Here we also have the title.
and some verdict. This is the reasoning. Let's pick another.
You see how many clips it did in just a short span of a time. And we also have kind of like a thumbnail that it pre -created. People want to find their passion, but you don't find it, you create it.
And you create it by getting good at something. And to get good at stuff, you start by doing something you suck at. Then you get good.
Then you like it. Then people ask. How you found your passion?
Answer. By starting when you sucked and not getting it.
Yeah, I hate the passion mantra. I really do. I mean, part of it was because I'll even play it out.
So I read all the same self -help books early on when I was in my entrepreneurial days and I was trying to find my passion and I was into fitness. And so I was deciding between fitness, test prep, because I was good at test prep.
I mean, starting is the perfect condition. Like whatever condition that you're in that you start was the perfect condition. So right there, we could see that it slightly cuts off at the very end.
I am going to click open and resolve. Let's see what happens.
By the way, this is a brand new feature. I have not tried this feature yet. Oh.
Holy, that is crazy. And then we can edit this ourselves if we want to. Although I think the timeline is kind of messed up.
What's the perfect condition? There we go. So if we cut that up, we can just do an in and out point.
And it's good. I had this idea about why you should stop taking advice from successful people. because most of their advice is not about what they did when they were at your stage.
It's about what they do now. And it's the same thing around pushing work -life balance. You know, what I've found after 50 years at Disney is that really the most important thing is, all right, well, how did you get to this stage?
What did you do when you were two years in? Anything that makes you less than $250 per hour for 30 minutes, I'm giving to somebody else. Trying to add on more to your schedule of all the things that you're not willing to sacrifice is the very reason you're not going to change it.
So right here. slightly it messes up so we'll do open and resolve it will recut the entire section oh it has the timeline now so right here we can see that it boom see the thing that i like most about doing this type of method instead of getting a straight mp4 is what if i wanted to extend this section right here Or I want to shorten it.
See, like if I want to make micro changes to this clip right here, I wouldn't be able to do that if this was already exported and rendered into an MP4. And so now what you're going to see is how I set up everything, the thinking behind it. And this is not going to be a step -by -step tutorial on this is what you typed in.
This is what you click here. This is going to be a higher level broad concept of how this all came to be and how you can apply and create this type of system into your own cloud agent or CHPT agent, whichever agent you use. It doesn't matter.
First, I started with Dave Ramsey, then the diary of a CEO one. Then Chris Williamson, which is over four hours all by itself. And what I had Claude do was I asked it the same thing every time.
Go find the complete question and answer chunks and cut them. That was 160 candidates in the system. It kept 107.
Now, 159 of the 160, it pointed at real moment in the recording, not a timestamp it invented, which is 99 .4%. Every accepted clip came out with a proposed title and a quote pulled for the thumbnail. And the whole thing, it took all three videos about 52 minutes of machine time.
Now, I do want to be a little bit careful with that number because 52 minutes, it sounds like I'm racing an editor. I'm not. That's machine time.
The thing I'm actually saying is nobody is hand cutting seven hours and 37 minutes of podcast in one afternoon. But every one of those 107 is a proposal. Most of them don't need any more editing, so I review them on a webpage way faster than opening up a timeline for each one.
Now, if one looks wrong, that's what DaVinci Resolve is for. The system, it writes every clip in its own timeline, so I open that one, fix it, and re -export it. Nothing goes out until I've watched them all.
So how's it deciding any of this? Now, Hermosi, he says his MoreMosi channel, it posts about 20 times a day. But when I counted the uploads myself, some days it was actually closer to a 40.
So call it between 20 to 40 a day, depending on who's counting. So out of those 900 uploads, I only rebuilt the long form question and answer lane. A question, it comes in and answer comes out.
In the system, it decides whether that answer is complete enough to stand on its own. That constraint, it matters because if I let it solve, make content from anything, I could technically move the goalposts on myself. after the fact.
So I wrote down what a publishable answer actually means before I asked the agent to choose anything. There has to be a real question in it, and the answer has to actually finish. A quote that only sounds interesting because you yanked it out of context doesn't count.
And neither does something that starts halfway through a thought. Roughly 2 to 10 minutes with a beginning that tells you what problem is on the table and an ending that feels done. Now, that's the pattern I measured in this channel's content, where somebody calls into a live stream and asks a question and gets an answer back.
Every rule in there, it came off a single channel making one kind of video. That's all I'm claiming here. Now, the useful part is that I can write most of it down.
Duration is just a number. Completeness, it leaves marks you can look for because a question has an edge and so does an answer. Those are the kind of rules a system repeats without getting tired.
Whether an answer it deserves a viewer's attention is the harder one and that comes later. Now, the first thing that actually broke was transcription. These hour plus videos were too long to do in one go.
My laptop back here, it just couldn't handle it and it would crash. So the agents, it broke each video into smaller segments. It transcribed those locally and then reassembled the whole thing back together.
Every single word ends up with its own timestamp. across three videos that came out to 93 ,220 words. And I checked the last timestamp against how long the footage actually was to make sure that nothing fell off the end.
Once the transcript exists, I don't ask Claude to guess that clip starts at 37 minutes and 12 seconds. AI is really good at copying words. It's much less trustworthy when it invents a timestamp out of a long document.
So it picks two points. The first words of the question and the last words of the answer. Then a plain text search hunts for those exact words in the time transcripts and snaps the proposal onto the real timestamps.
Plod, it proposes, but it's the search that decides where those words actually occurred. And that separation is the whole thing. Because it means a plausible sounding timestamp doesn't get to become a cut just because it came back looking official.
After that, the selector is the only place in the system that makes an actual editorial judgment. It gets those time question and answer chunks plus the written standard. It has to explain why a segment fits the kind of clip this channel publishes.
An accepted segment is still only a proposal because it hasn't watched the clip and it hasn't heard whether the cut starts on a clean thought or ends on an awkward half sentence. Here's what the selector wrote down when it made that call, word for word. And I did not write a syllable of this.
So 227, it sits inside the 120 and 600 second band. Open parentheses, below P25, but well within band share, close parentheses, reads as one bounded question. Where do founders misuse AI?
With a closing verdict, open parentheses, not making you more money, close parentheses, satisfying, requires single question unit. Okay, so what does that mean in plain English? It means the clip it runs 2 minutes and 27 seconds, which sits inside the 2 to 10 minute window I gave it.
It found one question with a clean edge on both sides and an answer that actually lands on a verdict instead of trailing off. That's the agent checking its own work against the standard IRO. None of this editing logic appeared out of nowhere, by the way.
A first pass scale I already built could find where talking starts and stops. It keeps 5 frame handles, it measures silence, It gave me something editable.
This project, it picked those pieces up and gave them a different job. And that's the part I think gets missed when people say, just have the AI edit it. The useful system, it's a pile of small rules that already survive actual footage.
One giant prompt doesn't get you there. Once a clip survives the selector, the system, it times the in and out points around the talking. It measures the silence sitting inside the clip.
It keeps the handles. And it writes the range into resolve. And this is where the system genuinely beats me.
Because cutting has rules I can describe out loud. Where does the question start? Where does the answer land?
How much dead air is in the middle? And did the range actually get written into the project file and saved? So comparing humans and agents, where does the agent win?
Anywhere I can write the rule down and repeat it. Matching words to time, tightening edges, cutting silence, building timelines. Doing the same operation across 100 clips without getting bored somewhere around clip 30.
At this volume, trusting the agent to cut is okay. I did find some minor mistakes and that's okay too. It doesn't have to be perfect.
The goal for this mass amount of volume isn't perfection. We're going for a volume game here. Now where it fell short on these three, the rule this thing runs on came from Hermosi's livestream lane where callers phone in, ask a question, and get an answer.
That's what the rule file was written for. And it says so in plain text. Every clip is built around a single caller's question.
The selector, it carries the same framing. So when I pointed it at a podcast, it went looking for a caller -shaped question on a show that doesn't have callers. And you can see it in the rejections.
Here's a piece of one exactly as I wrote it. A host -led aside writing on the previous answer's unicorn metaphor. Not an independent caller question.
Then there's the link floor. Same file says a published video clips run roughly two to 10 minutes because that's what the live stream lane produced. Out of the ones it threw out across these three, 41 were shorter than two minutes.
Conversational podcasts, it typically move faster than that. So a host and a guest trade quick exchanges back and forth in the floor. It has no idea.
So I left it alone. I could have turned the rule files until they separate. Look.
prettier on podcasts, but then I'm fitting the rules to three videos instead of telling you what the system actually did. Now, before I go on, this one's from earlier, not from these three. Back when I was building the system on a live stream, two words got stuck together in the transcript, basically one tiny glitch.
And because of that, the system, it lost almost nine minutes of a perfectly good answer. It just couldn't find it. I fixed it, ran it again, and the second attempt, it found way more of the recording.
But all that measures is how much of the recording the system could find at all, not whether its pics were any good. Still, on the earlier build, and this is the one that made me feel a little stupid, this next part is a reconstruction. I didn't have a camera running when it happened, so I opened a timeline, all 12 of them, and there were no clips inside any of them.
It was basically empty. Resolve had come back saying everyone succeeded, and the placements were never saved, and Resolve's default timecode put the clips an hour before the visible start.
And this isn't necessarily a problem with Claude, but a problem with DaVinci in the fact that DaVinci's timestamp or default timecode always starts at one zero zero zero zero zero zero instead of zero zero zero zero zero zero zero. Okay, that was a handful. So Claude by default, it starts everything at the timestamp zero, not one.
So the fix had nothing to do with the model. It saved the project, it set the timeline start to zero, then it read it back. A tool telling you it succeeded is not proof that a human can see the edit.
Okay, so where does an editor still win? I would say packaging. A system can surface a moment and then write a title that describes the topic without understanding why anybody would click on it.
A broken timestamp is wrong in a way you can improve. A title is a taste question. And the only honest judge of that is what real viewers...
do with it. So the agent hands me the option and I decide. What's still missing is the part that learns from publishing.
Although I can say that this is possible, I haven't automated publishing here and nothing here proves this system can't improve itself yet either. So you don't have to hand over your whole process over to a machine. I didn't.
Not one of those 107 goes anywhere until I've watched it myself. But the parts of the job where I can actually write the rule down. That's the stuff I stopped doing by hand.
If you can describe the step out loud, you can probably stop doing it yourself. Everything I just walked through, the scripts and the rule file it reads, that repo is linked in the description. It's not a download and go type of thing.
There is a setup and the README does not sugarcoat it. Only the step that writes into Resolve, it needs Resolve installed. Everything else you can run without it.
So go build it if you want. And if you want to talk through what this looks like on your own footage, I do a 45 minute research call.
Links in the description as well. If you want to see what happened when I put the agent up against me on the same footage, that one's linked below as well.
The Hook
The bait, then the rug-pull.
Andy Diep opens with the finished product: a wall of automatically cut podcast clips, each with a title, thumbnail, and a button that drops straight into DaVinci Resolve. Then he rewinds to show how a Claude Code system reverse-engineered from one creator's 900 uploads built it, and where those borrowed rules held up against podcasts they were never designed for.
Named ideas worth stealing.
10:16model
FAQ Engine selection policy
- one caller question = one video
- 2-10 minute length band (120-600s)
- require_single_question_unit
- closing verdict required, not a trailing-off answer
The written rule file the selector checks every candidate clip against before accepting it, reverse-engineered from what actually got published on the source channel.
Steal forany long-form-to-clips automation pipeline
12:02concept
Timestamp snapping via word search
Instead of trusting an AI-guessed cut point, the system proposes the literal first and last words of a segment, then a plain text search finds those exact words in the timed transcript to fix the real cut point.
Steal forany pipeline that turns free-text AI output into precise media timestamps
How they asked for the click.
VERBAL ASK
18:59link
“that repo is linked in the description... I do a 45 minute research call. Links in the description as well.”
soft, non-salesy close after the technical payoff, a free repo plus an optional free consulting call, no product pitch
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
