Skip to content

Code skill ยท free

Animated Captions

Word-by-word animated captions for talking videos. Transcribes and renders locally, free.

A 10.5-second clip this skill captioned. The words were transcribed on our computer and each one lights up as it is spoken. Footage "Timelapse of Clouds over Bellevue Canyon" by Extemporalist (CC0, via Wikimedia Commons); voice-over written by us and spoken by the Kokoro-82M voice model (Apache-2.0).

Download the skill

Install it in Claude Code, Codex, Cursor and run it on your own computer.

Download .zip

Get it on GitHub

The open-source version, free to use and change.

Open on GitHub

Install

  1. Download the .zip above and unzip it.
  2. Put the animated-captions folder in your agent's skills folder, for example ~/.claude/skills/animated-captions for Claude Code.
  3. Ask your agent for the video. The skill's SKILL.md tells it what to collect and which script to run.

Inputs

  • Video (video)

    The clip to caption, any shape. Vertical 9:16 suits Shorts, Reels and TikTok. The sound is kept.

  • Transcript (text, optional)

    Optional. Without it the skill transcribes the speech on your computer. An SRT or VTT file also works, with approximate word timing.

  • Caption position (choice, optional)

    Where the caption block sits on the frame. Bottom suits most clips; pick middle or top if the bottom is busy.

  • Accent colour and words per line (text, optional)

    The colour of the spoken word (yellow by default) and at most how many words show at once (4 by default).

  • Start and end (number, optional)

    Caption only part of the clip, in seconds from the start of the file.

Made with

Engine
HyperFrames
Output
video
Cost per run
Free: renders on your computer

Most short-form video is watched with the sound off, so captions carry the message. The style that works on Shorts, Reels and TikTok shows a few words at a time and highlights the word being spoken.

This skill teaches your coding agent to build exactly that. It transcribes the speech on your computer, splits the words into short lines at sentence ends, commas and pauses, and renders the captions over your clip with HyperFrames.

How it works

  1. Give the agent your clip, and optionally which part to caption.
  2. Pick where the captions sit, the highlight colour and how many words show at once.
  3. The agent transcribes the speech, shows you the caption lines to check, and renders a captioned MP4 with the sound kept.

Because the captions are code, fixing a misheard word means editing one word and rendering again.

Questions

Does the skill transcribe the audio for me?
Yes. It runs a free speech model on your own computer and times every word. The model is installed once with one command and is about 640 MB. You can also give it an SRT or VTT file you already have.
Is there an API cost?
No. Transcription and rendering both run on your computer with HyperFrames, so there is no per-video charge.
Does it keep the sound and the shape of my video?
Yes. The audio is kept, and a vertical clip stays vertical, including phone videos saved sideways with a rotation flag. Videos larger than 1920 pixels on the long side are scaled down.
What if a word is misheard?
The word timings are saved as a small JSON file. Fix the word there, point the skill at that file and render again.
How long does a render take?
Our 10.5-second 1080p sample rendered in about 40 seconds on a 4-core computer, after a 10-second transcription.

Sources