I recently used Gemini to help with some dashcam videos that weren't being saved properly. I was sure most of the data were there but the files wouldn't play in VLC or MPC, so I asked Gemini.
It suggested various things to try, and after pasting in the error messages each time it suggested more and more radical things. Eventually it suggested a program called Untrunc, where you give it a working video file as a reference and then the file that's faulty and as if by magic it worked!
Just wanted to mention this in case anyone else is struggling to get FFmpeg to repair a file.
Interesting - I took a look at how this works and why it needs a reference, and the answer is the usual one of the 'moov atom': a critical piece of metadata. Lots of programs output it at the end of the file, but that makes it vulnerable to truncation, and for streaming it's useful to move it to the beginning of the file. ffmpeg refers to this as "fast start".
Not OP, but the whole "AI using em-dashes when people writing would use commas" is very annoying given that the em-dash is readily available on my phone (hold on - to get one) and on my computer (compose - -), and I only started using it because people pointed out that commas are incorrect in this instance :(
FTR I don't, and have never used AI to write or do any "creative work", because it's not creativity if it doesn't come from you.
Yeah, same. To be clear I'm frustrated and annoyed that people are using LLMs to replace writing just... normal everyday things, that it is beneficial to think through and consider and form in your mind; it is likewise frustrating that immediately the first thing that anyone really did with AI was spam online submission forms for writers groups hard enough that they got taken offline.
Great work on this - I made a terminal command similar to this (llmpeg), and was actively trying to get exactly this working - a webasm compiled version of ffmpeg that could encode in the browser. I for the life of me couldn't get the provided examples on https://github.com/ffmpegwasm/ffmpeg.wasm to run.
Just for my own development curiosity, was there anything specific you had to do to get ffmpegwasm to work?
The creator of that repo has a blog with some posts on building a wasm version of FFmpeg, I mostly followed those: https://jeromewus.medium.com/
I also used some of their scripts for building the 3rd-party libraries.
It took a bit of trial and error to see which versions of the different libraries and build tools work together.
The Warp terminal[1] is excellent for this type of thing. In agent mode, you just describe what you want to have happen and it generates the proper command(s) (that you can approve before running).
I use it a lot to convert videos and turn a folder of tiff files into pngs at 1/2 size, etc. It's great at generating FFMEG commands and chaining the right tools together.
I was going to mention Warp here as well. It is fantastic when it comes to almost anything in the terminal. It has caused me to use the terminal a lot more on all of my computers because I don't have to spend a bunch of time poking around on Google to find the command to run.
I have used it for ffmpeg and then a lot of other slightly more complex commands. A recent one from the other day was gathering up all of the .epub documents in a directory tree, renaming them to the name of the directory they were in, and then placing them all in one single directory. That would have been a whole project for me, and Warp gave me the command with just that description. Any LLM interface would have done the same, but Warp just let me hit "Enter" and run it, no need to copy and paste.
IME as a dev, I got value from Aider using AI as a commodity to edit code. The sparkly buttons in AWS Cloudwatch log queries, Mongo Compass desktop app and other apps with AI features are hit or miss so far.
This is such a convenient tool for a casual user, and a great application of an LLM to a narrow task that probably couldn't be handled quite so easily everywhere. Also a great example of the emerging 'chat driven' UX trend, which I'm really liking.
I tried to implement something very similar recently, and had the hardest time getting the LLM to produce anything remotely resembling actual ffmpeg commands.
You were using a weak LLM then. The difference between one of the leading edge LLMs like Gemini 2.5 Pro, o3, or Claude 4 and an average LLM or one you can run on your typical PC/laptop is night and day.
I think you need to add some liberal filename handling. I have directories of videos generated by various AI video models, and they have spaces in the filenames, not just one, but the prompt to generate the video plus the major parameters are all the filename. They are long, pains in the ass to work with, but they are there. Would be nice if your tool could work with them.
Nice! Reminds me of the ooo-wow early LLM days where those little explain-Regexps one-off sites were popping up everywhere. Seems inevitable that the LLMs will slurp up this functionality natively as well
interesting, i created a video editor to generate ffmpeg commands because it was difficult to adjust ffmpeg commands on commandline https://newbeelearn.com/tools/videoeditor/ . This lets you tweak ffmpeg commands visually.
Honestly what a great application of LLMs. FFmpeg is a very powerful tool, and as with most powerful tools is very complicated to run correctly. Do the files get uploaded though? Or does it just grab the location on disk?
Therefore you lose any hardware acceleration I guess? I'd prefer to run an actual native ffmpeg on my machine, but of course then there's the security issue of copy-pasting a command line of "unknown" origin...
I recently used Gemini to help with some dashcam videos that weren't being saved properly. I was sure most of the data were there but the files wouldn't play in VLC or MPC, so I asked Gemini.
It suggested various things to try, and after pasting in the error messages each time it suggested more and more radical things. Eventually it suggested a program called Untrunc, where you give it a working video file as a reference and then the file that's faulty and as if by magic it worked!
Just wanted to mention this in case anyone else is struggling to get FFmpeg to repair a file.
Interesting - I took a look at how this works and why it needs a reference, and the answer is the usual one of the 'moov atom': a critical piece of metadata. Lots of programs output it at the end of the file, but that makes it vulnerable to truncation, and for streaming it's useful to move it to the beginning of the file. ffmpeg refers to this as "fast start".
That makes sense, thanks for that
To yes, and this —
I've started doing something similar on the command line with Claude Code that works incredibly well:
claude -p "use ffmpeg to convert myvideo.mov into an h264 video suitable for youtube upload" --dangerously-skip-permissions
Highly recommended! I use --dangerously-skip-permissions because I just want to set it and forget it and dont need to babysit the run.
`llm cmd` by simonw is even better for one-off commands. It spits out the invocation and you can repurpose as suits your needs. Thanks Simon!
simonw's llm is amazing. Basically disappointed me for trying to build any LLM powered cli tool
Em-dash alert.
Not OP, but the whole "AI using em-dashes when people writing would use commas" is very annoying given that the em-dash is readily available on my phone (hold on - to get one) and on my computer (compose - -), and I only started using it because people pointed out that commas are incorrect in this instance :(
FTR I don't, and have never used AI to write or do any "creative work", because it's not creativity if it doesn't come from you.
I also have regularly used hyphens, en dashes, and em dashes since college. Way preceding LLMs. Now I second guess my literary style.
Yeah, same. To be clear I'm frustrated and annoyed that people are using LLMs to replace writing just... normal everyday things, that it is beneficial to think through and consider and form in your mind; it is likewise frustrating that immediately the first thing that anyone really did with AI was spam online submission forms for writers groups hard enough that they got taken offline.
https://www.scottsmitelli.com/articles/em-dash-tool/
Great work on this - I made a terminal command similar to this (llmpeg), and was actively trying to get exactly this working - a webasm compiled version of ffmpeg that could encode in the browser. I for the life of me couldn't get the provided examples on https://github.com/ffmpegwasm/ffmpeg.wasm to run.
Just for my own development curiosity, was there anything specific you had to do to get ffmpegwasm to work?
I made a script of this sort too originally called "ffmpeg-english" last year. Interesting how people reacted to it:
https://news.ycombinator.com/item?id=40410637
I then made a more general version of it for all commands:
https://github.com/dheera/scripts/blob/master/helpme
Example usage:
The creator of that repo has a blog with some posts on building a wasm version of FFmpeg, I mostly followed those: https://jeromewus.medium.com/ I also used some of their scripts for building the 3rd-party libraries.
It took a bit of trial and error to see which versions of the different libraries and build tools work together.
I had the same problem - would love to see a working example.
I don’t know about the browser but I’ve been playing with a WASM build through this Go wrapper. Nice to not need CGO.
https://codeberg.org/gruf/go-ffmpreg
I'm a little confused as to how someone makes an "mpreg" joke in the project name, but censors "lame" in the README.
Lame could be considered an ableist slur, I guess?
The Warp terminal[1] is excellent for this type of thing. In agent mode, you just describe what you want to have happen and it generates the proper command(s) (that you can approve before running).
I use it a lot to convert videos and turn a folder of tiff files into pngs at 1/2 size, etc. It's great at generating FFMEG commands and chaining the right tools together.
[1] https://www.warp.dev
I was going to mention Warp here as well. It is fantastic when it comes to almost anything in the terminal. It has caused me to use the terminal a lot more on all of my computers because I don't have to spend a bunch of time poking around on Google to find the command to run.
I have used it for ffmpeg and then a lot of other slightly more complex commands. A recent one from the other day was gathering up all of the .epub documents in a directory tree, renaming them to the name of the directory they were in, and then placing them all in one single directory. That would have been a whole project for me, and Warp gave me the command with just that description. Any LLM interface would have done the same, but Warp just let me hit "Enter" and run it, no need to copy and paste.
Also in zed you can hit ctl-enter in terminal or in code and get the same
IME as a dev, I got value from Aider using AI as a commodity to edit code. The sparkly buttons in AWS Cloudwatch log queries, Mongo Compass desktop app and other apps with AI features are hit or miss so far.
This is such a convenient tool for a casual user, and a great application of an LLM to a narrow task that probably couldn't be handled quite so easily everywhere. Also a great example of the emerging 'chat driven' UX trend, which I'm really liking.
I think it's amusing how the most consistent usecase for "LLM for CLI" is to use ffmpeg.
Anyone that has built something with FFmpeg understands why this is at the top of HN.
I tried to implement something very similar recently, and had the hardest time getting the LLM to produce anything remotely resembling actual ffmpeg commands.
You were using a weak LLM then. The difference between one of the leading edge LLMs like Gemini 2.5 Pro, o3, or Claude 4 and an average LLM or one you can run on your typical PC/laptop is night and day.
Which model did you use? Claude has been spot on for all my needs.
load up the man pages into the context maybe?
for real? to me it works fine with just chatGPT (free)
I think you need to add some liberal filename handling. I have directories of videos generated by various AI video models, and they have spaces in the filenames, not just one, but the prompt to generate the video plus the major parameters are all the filename. They are long, pains in the ass to work with, but they are there. Would be nice if your tool could work with them.
I haven't needed it in a while, but I remember detox[1] being good for bulk renaming without me having to provide input.
[1]: https://github.com/dharple/detox
Nice! Reminds me of the ooo-wow early LLM days where those little explain-Regexps one-off sites were popping up everywhere. Seems inevitable that the LLMs will slurp up this functionality natively as well
Love this! Folks here should also check out https://wide.video if they're looking for a VERY full featured free in-browser video editor
interesting, i created a video editor to generate ffmpeg commands because it was difficult to adjust ffmpeg commands on commandline https://newbeelearn.com/tools/videoeditor/ . This lets you tweak ffmpeg commands visually.
a similar website https://ffmpeg.app/
I also recommend djcopley/ShellOracle if you are lazy like me and never remember the commands.
That's great. How could it work in browser? Is it a webassembly implementation of FFmpeg?
How do you know my name?!
It's been sent in the tracking cookie
This is a perfect use case for thin LLM wrappers!
Honestly what a great application of LLMs. FFmpeg is a very powerful tool, and as with most powerful tools is very complicated to run correctly. Do the files get uploaded though? Or does it just grab the location on disk?
Nothing is uploaded, everything runs locally.
This appears to use ffmpeg.wasm
Therefore you lose any hardware acceleration I guess? I'd prefer to run an actual native ffmpeg on my machine, but of course then there's the security issue of copy-pasting a command line of "unknown" origin...