Drop a Video into Claude Code, Type /effects, and 45 Effects Apply Themselves

Claude Code now has a video effects skill. Drag a video into the conversation, type /effects and the effect you want, and Claude renders yourvideo_effects.mp4. The audio stays as-is.

First, an overview of the effects:

One-line install:

git clone https://github.com/Vibha-Ramprakash/claude-video-effects ~/.claude/skills/effects

On Windows PowerShell, replace the path with $HOME\.claude\skills\effects, then restart Claude Code. The first run installs ffmpeg and Python and takes about 2 minutes. If you don't want to touch the terminal, just tell Claude Code to install the github.com/Vibha-Ramprakash/claude-video-effects skill into ~/.claude/skills/effects, and it will handle it itself.

Usage combines natural language with slash commands. For example:

/effects orbit from 3 to 6 seconds
/effects spotlight when I say "kill the lights"
/effects comment-box at the end, typing CLAUDE

You don't need to count frames yourself for timing. Say “when I say watch,” and Claude will find that second. Effects can be stacked and ordered in a single command. It checks the video frames first, and effects that might cover a face are placed behind the head. Output is 1080p; you can ask for 4K, but rendering is slower. Effects work best with vertical phone videos and a person in frame.

This skill includes 45 effects. These include brackets that lock onto the face, giant words dropping down behind the person, a checklist that ticks off items as you speak, a Follow button flipping to Following, orbit, spotlight, comment-box, todo, tier-list, phone-fan, and messages. If you want orbit, noise, phone-fan, and profile-wall to use your own reels, give Claude a folder of screenshots. For numbers in charts, counters, and titles, just tell it the actual values.

The author @vibha.in.progress has documented the limitations in the repo. These are starter versions and won't fully match the final videos in her reels. Her original videos also have sound effects and hand-tuned timing. The skill only produces drafts, so review them yourself before publishing. It doesn't add sound or subtitles; add those in CapCut. If the render is wrong, just tell Claude “the word is covering my face” or “it's too fast,” and it will revise the plan and re-render.

The effects overview is in the GitHub repo, with all 45 effects and their parameters. The author is @vibha.in.progress. Face tracking uses OpenCV YuNet, portrait segmentation uses MediaPipe Selfie Segmenter, and font and code licenses are in the repo. The original tweet is here.

Some commenters said that dropping in a video and saying “spotlight when I say kill the lights” is enough to get the effect—a pretty wild workflow. Some think the Follow button flipping to Following is a bit evil. Other developers asked whether face checks still work when orbit and locked-on are stacked, and whether the output is checked again after rendering. Another commenter mentioned that a render for a published video silently failed at 7.7 seconds and was only discovered via a 2fps contact sheet.

Automatically adding effects saves the time of opening an editing app and placing effects frame by frame. Reliability still depends on human review. Playing it through before publishing is faster than only reading logs.