How it works:

• inputs are analysed using ffprobe
• the prompt and analysis are passed to gpt-5 for generating an ffmpeg command
• the command is run, if it fails, the failure is passed back to gpt-5
• the model attempts up to 3 times