Flowly
Log inDownload
Use cases/Creative & Building

AI Video Editing via Chat

Edit videos by describing changes — trim, merge, add subtitles, color grade, crop to vertical. ffmpeg under the hood, no timeline, no GUI.

Creative & Buildingmedium~30m setup
Tools
execread_filewrite_file
Channels
desktop

For repetitive video work — trimming intros from a batch, adding captions, cropping to vertical for shorts, color-grading 10 clips identically — opening a timeline editor wastes hours. ffmpeg can do all of this from the command line; the agent translates your plain English into the right ffmpeg incantations.

What it does

  • Take a video file path; describe the change in plain language
  • Agent generates the ffmpeg command, runs it, returns the output path
  • Common operations: trim, merge, crop, scale, subtitles, audio swap, watermark, color adjust
  • Batch operations: "do this to every .mp4 in the folder"
  • Preserves originals (always writes to a new file)

What you'll need

  • ffmpeg installed (brew install ffmpeg on Mac, apt install ffmpeg on Ubuntu)
  • Exec tool with a working directory you trust
  • Whisper (optional — pip install openai-whisper for local caption generation)

Setup

1. Verify ffmpeg

bash
ffmpeg -version
# Should output ffmpeg version 4.x or higher

2. Set the editing convention

Send to Flowly
Remember the video editing protocol: When I describe a change to a video file, you: 1. Confirm the source path exists; if not, ask me to clarify 2. Generate the ffmpeg command for the operation I described 3. Show me the command before running, in a code block 4. After my OK, run it via exec 5. Output file convention: <original-name>.<operation>.<ext> (e.g., interview.mp4 → interview.trimmed.mp4) 6. Originals are NEVER overwritten. Use a fresh output path every time. 7. After ffmpeg finishes, run "ffprobe -v error -show_format <output>" to verify duration, bitrate, codec — show me the summary. 8. If ffmpeg returns a non-zero exit, show stderr, propose a fix. I'll typically describe operations like: - "trim the first 30 seconds from <file>" - "merge clips A, B, C into one in that order" - "crop <file> to 9:16 vertical, center the action" - "scale to 1080p" - "add captions from <file.srt> burned in" - "auto-caption (use whisper) and burn captions in" - "extract audio as mp3" - "speed up 1.5x with pitch correction" Avoid: - Lossless ops where possible (use -c copy for trim/concat — no re-encode) - Re-encoding when the user didn't ask for it - Codec changes unless explicitly requested

3. Try simple operations

"Trim the first 30 seconds from ~/Movies/interview.mp4"

Agent shows:

bash
ffmpeg -i ~/Movies/interview.mp4 -ss 30 -c copy ~/Movies/interview.trimmed.mp4

You approve, it runs, returns the output path + ffprobe summary.

4. Auto-captioning

"Add auto-generated captions to ~/Movies/talk.mp4 — burn them in, white text, bottom centered"

Agent runs whisper to extract a SRT, then ffmpeg to burn the captions:

bash
whisper ~/Movies/talk.mp4 --output_format srt --output_dir /tmp/
ffmpeg -i ~/Movies/talk.mp4 \
  -vf "subtitles=/tmp/talk.srt:force_style='Alignment=2,Fontname=Arial,FontSize=24,PrimaryColour=&Hffffff'" \
  ~/Movies/talk.captioned.mp4

Confirms with you between the two stages so you can review the .srt before burning.

5. Batch mode

"For every .mov in ~/Movies/2026-04/, convert to .mp4 with H.264, keep originals."

Agent generates a one-liner that loops, shows it, then runs.

Tips

  • Lossless trim/concat is fast. -c copy skips re-encoding. Use it whenever the operation doesn't change codec, resolution, or bitrate.
  • Preview before batch. A batch operation that's wrong wastes hours. Run on one file first, check the output, then unleash on the folder.
  • Keep captions separate first. Auto-generated captions need proofreading before burning. Have the agent output the .srt, you edit, then burn.
  • Vertical crops need framing. "Crop to 9:16" without framing guidance loses the action. Tell the agent which side to favour: "9:16 with the speaker centered" works.
  • Check disk space first. Re-encoding multi-GB files temporarily doubles disk use. The agent should df -h before batch ops.
  • Audio sync issues. When concat'ing clips with different audio configurations, audio drifts. Tell the agent to re-encode audio in those cases (-c:a aac).
  • This is for utility, not artistry. For colour grading, transitions, motion graphics, real video editors are still better. Use the agent for the boring parts.