Best AI Prompts for Video Editing Scripts (2026)

Prompts for video editing#

Act as a video processing engineer. Write an FFmpeg shell script that batch-processes every file in [DESCRIBE YOUR INPUT FOLDER AND FORMATS]. Normalise loudness to [YOUR TARGET, e.g. -16 LUFS], optionally apply the colour LUT at [PATH TO YOUR .cube FILE], and export H.264 MP4 at [YOUR RESOLUTION AND FRAME RATE]. Write to a separate output folder, never overwrite originals, skip files that already have output, and print one line per file with success or failure. Show a dry-run mode first.

Act as a Python developer who knows FFmpeg. Write a script that takes [PATH TO A SOURCE VIDEO], detects scene changes, saves one thumbnail per scene and a short preview GIF, and writes a CSV of scene start times. Call FFmpeg through subprocess with argument lists, explain which detection threshold you chose and how I should tune it, and say what to do when no scene is found.

Act as a cloud engineer. Design a transcoding and watermarking job for files uploaded to [YOUR STORAGE SERVICE] that runs on [YOUR CLOUD SERVICE]. Show the FFmpeg command, the container setup, retry rules for failed jobs, how failed files are reported, and what it costs to run as a function of file length. List any assumptions about timeouts and memory.

Act as a motion graphics developer. Write a script that overlays captions from [PASTE YOUR SRT OR JSON CAPTION FILE] and the logo at [PATH TO YOUR LOGO FILE] onto [PATH TO A CLIP]. Use [REMOTION OR FFMPEG, PICK ONE] and say which version of each tool you assume. Keep captions inside safe margins for [YOUR TARGET ASPECT RATIO] and explain how I swap the font and position.

Act as a media asset specialist. Write a script that walks [PATH TO A VIDEO LIBRARY], reads each file's duration, resolution, codecs and creation date with ffprobe, and writes one JSON record per file with the fields [LIST THE TAGS YOU WANT, e.g. source, rights holder, licence]. Flag files that cannot be read, and never modify or move a source file.

What to add to any video-editing prompt#

• The exact tool and version. Say "FFmpeg 7.1 on macOS" or "Remotion 4". Google's Gemini guide advises clear and specific instructions and stating constraints, and Anthropic says to be explicit about the output you want.
• A sample of the input. Paste the output of ffprobe for one file (codec, resolution, frame rate, audio channels). The model then stops guessing the stream layout.
• A cut-accuracy requirement. Say whether cuts must be frame-accurate. FFmpeg documents that in most formats it is not possible to seek exactly, and that stream copy cannot be combined with filters.
• Safety rules for files. "Never overwrite inputs", "write to a new folder", "print the command before running it". FFmpeg's -y overwrites outputs without asking and -n refuses to.
• The output format of the answer. Anthropic and Google both advise specifying the output format; ask for "one script plus a short list of assumptions".
• A request to test. OpenAI's guide says to instruct the model to test changes; ask for a test command that runs on a 10-second clip.

A worked example#

Act as a video processing engineer. Write an FFmpeg shell script that batch-processes every .mov file in ./raw. The files are 4K 29.97 fps H.264 with stereo AAC audio (ffprobe output pasted below). Normalise loudness with the loudnorm filter in two passes, scale to 1920x1080, and export H.264 MP4 with AAC audio to ./export. Never overwrite originals, skip files that already exist in ./export, and print one line per file. Add a dry-run flag that prints the commands without running them. Finish with a test command for a 10-second clip.

ffprobe output: [PASTE YOUR FFPROBE OUTPUT]

A good result for this task must contain:

• A loop that quotes file names, so spaces in names do not break the script.
• Two loudnorm passes, with the measured values from pass one fed into pass two, as the filter's linear mode requires.
• Output to a separate folder, with -n or an existence check instead of a blanket -y.
• A dry-run mode and a short-clip test command.
• Explicit -map choices if the files have more than one audio or subtitle stream.

Check rights and licences too. Footage, music, fonts, LUTs and logos you process keep their own licences. FFmpeg's legal page says that GPL parts, notably libx264, make the GPL apply to all of FFmpeg, so a build you distribute with them has obligations. A cloud job that processes other people's uploads needs their permission and your own retention rules.

For generating clips from text, see our text-to-video prompt guide. The full prompt library is on our prompt library page.

Limits of this guide#

We did not run these prompts or the commands they produce, and we did not measure which model writes the most reliable FFmpeg. We read the FFmpeg main documentation, its filters page (loudnorm), its legal page and three vendor prompting guides. We could not confirm details of the select scene-detection and subtitles filters or of Remotion's caption API from the pages we read, so test those parts on your own files and versions. Nothing here covers tools for editing inside an app such as Premiere or DaVinci Resolve.