Edited short
A carefully edited short: grade, captions or keywords, title behind the head
Polish one real talking-head clip of 1-5 minutes into a 40-90 second vertical video (1080x1920) without re-shooting: word-level cut, natural grade, keywords and captions.
Important: this is a skill for terminal AI coding agents (Claude Code, Codex) that runs on your own computer. It is not a prompt for the ChatGPT website. Check the requirements below before buying.
What it makes
You filmed yourself on a phone. The take has fillers, restarts and pauses, the colour is flat, and adding text by hand takes longer than recording. This look does the edit:
- Word-level cut: fillers, restarts and dead air removed, then every join re-transcribed and checked.
- Natural grade:
tu-nhienby default, or a warm cinematic preset for real estate and premium products. No skin smoothing. - Face-aware layout: text never touches the mouth or chin, and a round speaker bubble shrinks over maps, demos or terminals.
- On screen: keyword overlays, captions, held cards, real-data chips and title effects, including a title behind the speaker's head.
- Optional Before/After hook: the raw clip and the cut side by side in the first seconds.
Two looks: clean-lime (default) and bds-neon-purple for real estate or premium products.
How it works
Open your agent next to your clip and ask, for example:
Dựng clip này thành short dọc khoảng 50 giây, có phụ đề và từ khóaThe agent asks at most four things (target length, look, Before/After hook, which facts and assets the video may show), then transcribes, cuts, checks the joins, grades, builds the edit and renders a quick draft for review before the full-quality render. Every number, name and claim on screen is spoken in the kept part of the clip or comes from a fact you supplied with a source.
What you get
- One file,
renders/<slug>-9x16-phu-de.mp4, plus a QA report (qa/report.md). A second version without captions only when you ask. - Blocking layout checks: a broken layout stops the build instead of reaching the render.
- A path for long talks of 10-30 minutes: cleaned up and split into chapters, or cut into shorts.
What you need
- Claude Code or Codex.
- ffmpeg and Node.js 18 or newer (HyperFrames runs through
reelcrew hyperframes, elsenpx hyperframes). - A Python environment with
opencv-python-headlessandnumpy; the setup commands are in the skill. - A
SONIOX_API_KEYfor speech-to-text.
Install and make your first video
Copy the skill folder into your skills folder:
~/.claude/skills/talkinghead-edit-video/For Codex, use ~/.agents/skills/talkinghead-edit-video/. Then open the agent in a project folder and point it at your clip.
Who it is for
Polishing one real talking-head clip of a founder, agent, coach or teacher: word-level cut, grade, keywords, captions and a round bubble over maps or demos.
Not a fit if you want
- A faceless video or a text-to-speech voice.
- A line-diagram explainer.
Frequently asked questions
Will it add numbers or claims I did not say?
The skill requires anything on screen to be spoken in the kept clip, to be a fact you supplied with a source, or to be labelled as approximate or illustrative. Still read the video before you post.
Does it show where I filmed?
No. GPS, coordinates and the shoot location are never read or shown. The local shoot time is allowed.
Can I use the videos commercially?
Videos you render are yours for any commercial use. You may not resell, redistribute or share the skill files.
More looks from Talking Head
Full set: Talking Head
3 looks

Collage
The speaker as a paper cutout, with text slamming in on each word

Diagram
A line diagram growing beside the speaker's face bubble
Included in ReelCrewHighlight
A highlight short with motion-fx design




