Sabitlenmiş Tweet
sam
2.9K posts
sam
@samgrows
19, founder @algrow_online + economics undergrad
united kingdom Katılım Eylül 2024
591 Takip Edilen7.3K Takipçiler
algrow.online/uptime if u also want to monitor the stats of our voice generation models
English
@borjitaea @saimagnate what i mean by highly skilled workers is people who are actually in the YT field and understand what a good output looks like
not really highly skilled devs but QA checkers e.g. faceless creators or video editors reviewing graphics
hopefully u get what i mean
English
@saimagnate Not as simple as you think to make a complex product of this, but neither as "highly skilled" as Sam states
Bottlenecks here aren't about skill IMO
English
I retardmaxxed and vibe coded my own autonomous ai video editor in 3 days, and here’s how you can too
It shouldn’t be hard and expensive at all for anyone willing (considering it’s based on your own niche and isn’t for commercial purposes).
Cost per 15 min video: ~ $1.75 (electricity usage included)
Final video consists of:
Images (google images)
Clips (YouTube)
Zero stock footage
Niche: Sports and celeb documentaries
*BUILT WITH CODEX
Training data for Claude:
1) My own videos
Permanent assets:
Transition pack
Sfx pack
4-5 songs from yt audio library
Multiple Animated backgrounds
- APIs used:
1) Brave search
2) Claude
3) Gemini
4) Remotion
5) yt-dlp
Rendering:
- Running locally on my PC
Here’s the full pipeline:
- Script + VO is fed to the tool
- Claude (sonnet 4.6) reads the script and develops a scene plan
Scene plan includes:
1) Number of beats (visual changes on screen)
2) Ratio of clips + images (it’s trained on my own videos as mentioned earlier)
3) search terms for clips and images to be used (images are searched on google images through brave search and clips are sourced through yt-dlp from YouTube)
4) text shown on title cards and lower thirds
5) number of transitions, which ones and where to use them
4) text and info for motion graphics (graphs etc) and where to insert them
- Gemini (2.5 flash):
1) Quality checks all sourced visuals (images + clips) and rejects/accepts based on the guidelines I have fed it
2) Upscales images that are essential to the script but are low quality
- Claude (sonnet 4.6) again:
The main brain behind stitching everything together and coordinating between all the modules of the tool, now it takes all the finalised assets and runs them through remotion for rendering
In the development phase, codex itself was the brain behind coordination across all the modules it built. Once I got it packaged into a desktop app, the brain has since now been Claude.
For packaging it into a desktop app, I used:
Claude design for frontend
Electron for the app itself (all handled by codex)
I would certainly say it’s not super simple but not super hard to figure out either
* The example video below is from very early prototype phase and shows what I was able to achieve by day 2, it doesn’t show motion graphics but the current version does have them.
sam@samgrows
been thinking about what it actually takes to build ur own Vidrush style tool. heres where i'd focus if i was starting from scratch today i do think anyone can build a Vidrush competitor. but from what i've seen, you need a lot of capital and highly skilled team members to actually pull it off doing this solo for a lot of users prob isnt possible. could be wrong but theres a lot that goes into a website like this and vibe coding isnt gonna cut it for a full scale app. if i was in the position most ppl are in trying to build their own Vidrush, i'd probably tackle one format at a time. main thing Vidrush excels at from what i can see is top 10 listicle styles. top food brands ppl eat, stuff like that and ur actually in a decent spot rn because when Vidrush was building up they didnt really have many examples to go off. now you have pure Vidrush channels to feed to ur AI but i dont think thats enough on its own if i was building my own, the main thing i'd try to focus on is getting cost of production low and margins high, and pumping videos out as quick and as fast as you can heres what i think is really key and i could genuinely be wrong here: retention graphs once you have a retention graph and ur feeding that back to AI, its honestly one of the best things you can get imo. retention + comments are just so OP for iterating ur own local model because u literally have raw feedback from viewers even if u know friends with vidrush channels, ask them for screenshots of their retention graphs. get AI to map it second by second and analyze what went wrong. stuff like that. footage sourcing is the other main thing. for motion graphics, remotion seems to be the move from what im seeing. today they dropped a really good update its called HTML to frames. way easier when making motion graphics because before,u had to screenshot and recreate in JSX. now HTML elements can be treated like any other element and warp/mass transformed. makes it easier for everyone one direction that could work: if u find a Vidrush channel popping off, u can train ur system on that specific channel. hopefully the vidrush guys dont outscale u lol to top it off, i prob wouldnt go into this commercially unless u have a lot of capital or co founders who are actually good devs and i'm not bashing anyone trying it at all, got a close friend doing pretty decent with one and his cost of production is fairly low. if u think u can do it then go for it, im just one perspective honest advice from where im sitting: if ur trying to build a Vidrush to compete, prob better to do it locally. dont try to scale up to a full app for users. just make it for yourself or maybe one other person. seems like u run into way less issues that way but yeah thats just my take, could be missing stuff, and its only gonna get easier as AI progresses
English
@JamesNwobu Yes, we noticed it last minute but left it in for possible engagement
English
been thinking about what it actually takes to build ur own Vidrush style tool.
heres where i'd focus if i was starting from scratch today
i do think anyone can build a Vidrush competitor. but from what i've seen, you need a lot of capital and highly skilled team members to actually pull it off
doing this solo for a lot of users prob isnt possible. could be wrong but theres a lot that goes into a website like this and vibe coding isnt gonna cut it for a full scale app.
if i was in the position most ppl are in trying to build their own Vidrush, i'd probably tackle one format at a time.
main thing Vidrush excels at from what i can see is top 10 listicle styles. top food brands ppl eat, stuff like that
and ur actually in a decent spot rn because when Vidrush was building up they didnt really have many examples to go off.
now you have pure Vidrush channels to feed to ur AI
but i dont think thats enough on its own
if i was building my own, the main thing i'd try to focus on is getting cost of production low and margins high, and pumping videos out as quick and as fast as you can
heres what i think is really key and i could genuinely be wrong here: retention graphs
once you have a retention graph and ur feeding that back to AI, its honestly one of the best things you can get imo. retention + comments are just so OP for iterating ur own local model because u literally have raw feedback from viewers
even if u know friends with vidrush channels, ask them for screenshots of their retention graphs. get AI to map it second by second and analyze what went wrong. stuff like that.
footage sourcing is the other main thing. for motion graphics,
remotion seems to be the move from what im seeing.
today they dropped a really good update
its called HTML to frames.
way easier when making motion graphics because before,u had to screenshot and recreate in JSX.
now HTML elements can be treated like any other element and warp/mass transformed. makes it easier for everyone
one direction that could work: if u find a Vidrush channel popping off, u can train ur system on that specific channel. hopefully the vidrush guys dont outscale u lol
to top it off, i prob wouldnt go into this commercially unless u have a lot of capital or co founders who are actually good devs
and i'm not bashing anyone trying it at all, got a close friend doing pretty decent with one and his cost of production is fairly low.
if u think u can do it then go for it, im just one perspective
honest advice from where im sitting: if ur trying to build a Vidrush to compete, prob better to do it locally.
dont try to scale up to a full app for users. just make it for yourself or maybe one other person. seems like u run into way less issues that way
but yeah thats just my take, could be missing stuff, and its only gonna get easier as AI progresses
sam@samgrows
For everyone vibe coding their own personal Vidrush, unfortunately it goes a lot further than just Pexels and Pixabay
English
Took at shot at him??
I didnt even know he tweeted anything out haha 😂 I'm just seeing it now and good on him
The reason I tweeted is cause I got a bunch of dms about ppl asking about how to build their own Vidrush and I just wanted to let ppl know its not as easy as it seems and u gotta do more digging
Good stuff on u @saimagnate and glad to see what u cook up
English
@saimagnate @ahmdrafqq @samgrows Not saying he’s a bad person because I don’t know him but he’s clearly talking about you and he threw a shot at Vidrush when they announce their MCP. We need to stop the weird vibes in this community just build, get better and profit. You’re the only person I see helping for free
English
Insane amount of uses since release
Feedback from u guys has been insane and this feature is still unlimited usage for now
So go give it a go
sam@samgrows
BREAKING: Claude can now watch videos for you! You can plug in a YouTube, Tiktok or Instagram URL and let Claude break it down for you Live now inside of Algrow
English
@samgrows That's a massive time-saver. Empowering creators to focus on what they do best. Love to see it!
English
@abg_creeptoe We process it on our end in depth and Claude sees the results etc and parses it
English
@samgrows How does it “sees” the video?
Does it take couple of screenshot every seconds?
English
@dherealmark I do some development work
But the team consists of
- lead developer
- 2 growth marketers
- customer support
-in house motion designer
English
@samgrows So, this is a whole team building this? Not just you?
English
@MarcioClutch @thebadassdev it sees the video directly and can hear the audio. we process it on our backend
English
@samgrows @thebadassdev Does it view a sequence of images and tries to understand movement, or watched video directly without intermediaries
English
