Manoj | The AI Dude (@Manojkumar2257) on X

What it claims

heditor is an MIT-licensed agent skill from pycodebr that turns a raw talking-head recording, such as a class, a livestream or a podcast, into an edited clip. It runs inside Claude Code, Codex, Cursor or OpenCode. The post, by Manoj (@Manojkumar2257) on X, says the tool picks one complete 30-70 second thought "without inventing speech." The edited audio is built only from what the speaker actually said.

How it works

The post lists five stages:

1. Transcribes with word-level timings using faster-whisper, then selects one complete 30-70 second thought.
2. Cuts frame-accurately, shortens pauses and cleans the voice.
3. Designs motion graphics for that specific clip in HTML and GSAP.
4. Mixes with the voice prioritised, at -14 LUFS.
5. Runs QA, which has four checks: vision-reviewed previews, a loudness check, a fresh transcript of the final audio, and a 40-frame contact sheet.

The re-transcription of the final audio lets the agent compare the output against the intended cut.

The 42 named styles include podcast talk show, retro RPG HUD, courtroom sketch, chalkboard lesson and trailer. There is also a plain cut with captions only. Long-form mode removes dead pauses and false starts, and adds chapters, SRT/VTT subtitles and lower thirds.

How to use it

Install the skill globally:

npx skills add pycodebr/heditor -g

Then run the repository's setup script. After that, prompt the agent:

edit this video in creative mode

The repository is linked as github.com/pycodebr/hedit… (the URL is truncated in the post).

Limits

The source is a promotional X post. It gives no system requirements, no GPU or runtime needs for faster-whisper, no setup script name, no processing times and no output examples. It does not say whether non-English speech is supported. It does not name the trigger phrase for long-form mode. The claim that one recording yields "a week of Reels" sellable as a monthly package is the poster's opinion, not a measured result.