AI Video Editing Prompts: 10 You Can Copy for Claude Code

What these AI video editing prompts do, and what they need

These AI video editing prompts are short instructions for one job each. Paste one into Claude Code in the folder that holds your clip, and Claude writes and runs the ffmpeg or Python commands that do it. They are written for talking-head video: one person, one camera, speech that drives the edit.

Four rules apply to all ten:

• Make the transcript first. Most prompts read words.json, a word-by-word transcript with timings. The one-time setup prompt creates it with faster-whisper.
• Keep the original. Each prompt writes a new file. Swap input.mp4 for the latest file in your chain as you go.
• Cut before you decorate. A cut shifts every later timestamp, so run the clean-up prompts first, then ask Claude to refresh words.json.
• One prompt per message. Square brackets mark values you fill in.

For the anatomy behind prompts like these, read how to write a video prompt Claude can follow.

Clean-up prompts: silences, filler words and loudness

Clean-up comes first because every later edit depends on the timeline.

1. Trim silences

What it does: shortens long pauses without clipping words. What you need: the clip and words.json.

Tighten the pauses in input.mp4.
1. Run ffmpeg's silencedetect with noise=[-35dB] and d=[0.5]; list each
   silence with its start, end and length.
2. Shorten each silence to 0.25 s. Never cut inside a word in words.json.
3. Show me the cut list and wait. After I approve, render trimmed.mp4.

silencedetect logs audio that stays at or below a noise level for a minimum duration. Its defaults, -60 dB and 2 seconds, suit long gaps rather than speech, so the prompt sets both.

2. Flag filler words

What it does: lists likely fillers so you choose what goes. What you need: the clip and words.json.

Find filler words in input.mp4: um, uh, and "you know" or "like" as filler.
Also list every gap over 0.3 s between words in words.json where
silencedetect finds no silence, as "unlabelled sound".
Make a table: time, word, and the sentence around it. Cut nothing yet.

The gap check matters. A study of speech recognition on non-native English found that Whisper, the model faster-whisper reimplements, removed most fillers from its transcripts unless prompted, and that prompting could make it add words never said.

3. Even out the loudness

What it does: brings the whole video to one loudness target. What you need: the edited clip and a target.

Normalise trimmed.mp4 with ffmpeg's loudnorm filter in two passes.
Pass 1: measure with I=[TARGET], TP=-1.5, LRA=11, print_format=json.
Pass 2: same targets, the four measured_ values from pass 1, linear=true.
Copy the video stream, audio AAC at 48 kHz, save loud.mp4. Report the
before and after. If loudnorm falls back to dynamic mode, tell me why.

loudnorm is ffmpeg’s EBU R128 normaliser. Linear mode reverts to dynamic if the target range is below the source’s or the gain would push peaks past the limit, and dynamic mode upsamples to 192 kHz, hence the fixed output rate. Targets run from -70 to -5, default -24; if unsure, match a finished video of yours.

On-screen prompts: captions, punch-in zooms and a title card

These three prompts add what the viewer reads and the movement that marks key moments.

4. Word-timed captions

What it does: captions that appear on each spoken word. What you need: words.json, a font and an accent colour.

Make captions for input.mp4 from words.json.
- 1 to 3 words per caption; never across a pause over 0.3 s.
- Each word appears at its own start time; the word being spoken is
  [ACCENT HEX], the rest white. [FONT], [SIZE] px, lower third, off my face.
- First correct these spellings in words.json: [NAMES AND TERMS].
Write captions.ass, render the first 10 s at 540×960, then wait.

ffmpeg’s subtitles filter draws subtitle files with the libass library and exists only in builds compiled with it. Without it, Claude can draw the captions in Python.

5. Punch-in zooms on key words

What it does: a quick push-in on the words you stress. What you need: words.json and room around you in the frame.

Add punch-in zooms to input.mp4.
1. From words.json, suggest [N] words I probably stress, with times. Wait.
2. On each approved word: scale to 115% over 0.15 s, centred on my face,
   hold to the end of the sentence, ease back over 0.3 s.
3. At least 2 s between zooms. Render a 540×960 preview.

Claude picks words from the text, not your delivery, so edit its list. The free Digital Zoom prompt adds two-colour captions.

6. Title card

What it does: an opening card over the first two seconds. What you need: a title, a font and two colours.

Add a title card over the first 2.0 s of input.mp4: "[TITLE]" in
[FONT] Bold, [SIZE] px, [TEXT HEX] on [BACKGROUND HEX].
Fade in over 0.3 s, hold, wipe up over 0.4 s to reveal the video.
Keep my audio from 0.0 s, and the size and fps of input.mp4.
Show me one still of the card first.

Structure and delivery prompts: B-roll, chapters, exports and QA

The last four prompts plan cutaways, mark chapters, make each format and check the finished file.

7. B-roll placement notes

What it does: a timed list of cutaways, without touching the video. What you need: words.json.

Plan B-roll for input.mp4 from words.json. Do not edit the video.
For each moment: start, end, the sentence I am saying, what to show, and
full frame or corner. About one every [12] s. Save broll_notes.md.

Claude Code places footage you supply; it does not film anything.

8. Chapter markers

What it does: chapter titles for the description and the file. What you need: words.json of a long video.

Split input.mp4 into chapters using words.json: first at 00:00, at least
three, each at least 10 s, titles of 2 to 5 words in my own wording.
Write chapters.txt ("00:00 Title" per line) and chapters.ffmeta in
ffmpeg's FFMETADATA format, then add them to a copy without re-encoding.

The rules come from YouTube’s help page on chapters. ffmpeg’s metadata format stores each chapter as a [CHAPTER] section with a start, an end and a title.

9. Export 9:16 and 16:9

What it does: one master, two shapes. What you need: a 1920×1080 master without burned-in captions.

Make two exports from master.mp4 (1920×1080):
1. landscape.mp4 at 1920×1080.
2. vertical.mp4 at 1080×1920: a 9:16 crop that follows my face, smoothed
   to move at most [8] px per frame. Add captions after cropping.
First show me four frames with the crop box drawn on the landscape image.

A 9:16 window from a 1080-pixel-high frame is about 608 pixels wide, so the vertical version is upscaled. For Shorts, YouTube’s upload page asks for square or vertical video of up to 3 minutes.

10. QA check on frames

What it does: a list of problems with timestamps; it fixes nothing. What you need: the final render.

Check final.mp4. Change nothing.
1. Report resolution, fps, duration and loudness (loudnorm summary).
2. Export a frame every 2 s and a 4x4 contact sheet (ffmpeg tile filter).
3. Check captions inside the frame and off my face, spelling, black or
   frozen frames, and zooms that crop my head.
4. List each problem with its time and frame file.

Claude Code’s file reader shows Claude PNG, JPEG, GIF and WebP images, so it checks stills, not motion or lip sync. Watch the file once yourself.

Prompts, skills or an editing app: which fits the job

A copy-paste prompt fits a job you can describe exactly and repeat now and then. Here is how Claude Code editing prompts compare.

If you paste the same AI video editing prompts every week, save them as a skill. How to automate video editing with Claude Code covers that step.

Next: the free effect prompts and the setup

The quickest next step is the setup prompt, then one effect. Pasted once per video folder, the setup installs ffmpeg, writes words.json, cuts a person mask for every frame, adds shared helpers and a caption style, and exports a test clip.

For layered Claude Code video effects, the 17 free effect prompts work as a free video effects pack in five groups: depth and layers (six, such as Orbit), screens and UI (four), camera moves (three), text and graphics (two) and hand-made looks (two). With several working, the Combine & Fix workflow puts them on one timeline and saves the setup as a skill.