Communitygithub.com

talking-head-reel

Edit a talking-head clip into a 9:16 reel or ad - a slow-zoom hook with big two-line captions, punch-ins at every cut, a small set of scene types on the lines that need them, and captions under the face. Use for "edit this talking-head video", "make a reel from this clip", "turn this founder video into a reel or ad".

What is talking-head-reel?

talking-head-reel is a Claude Code agent skill that edit a talking-head clip into a 9:16 reel or ad - a slow-zoom hook with big two-line captions, punch-ins at every cut, a small set of scene types on the lines that need them, and captions under the face. Use for "edit this talking-head video", "make a reel from this clip", "turn this founder video into a reel or ad".

Works with~Claude Code~Codex CLI~Cursor
npx skills add https://github.com/justlowys/claude-video-editing-skills/tree/HEAD/skills/talking-head-reel

Ask in your favorite AI

Open a new chat with this agent skill pre-loaded.

Documentation

Talking-head reel

Read video-edit-rules first. Its rules apply to every step here.

The goal: the speaker stays the star, and the words are easy to read with the sound off. The test: a viewer reads the hook in the first second, the face is never covered, and the speaker is on screen for at least 85% of an organic reel.

For a paid ad, finish here, then run ad-video-editor for hooks, sizes and the ad checks.

Footage that works

  • One person talking to camera, 20 to 90 s.
  • Horizontal 4K is best. The 9:16 crop of 4K is 1215x2160, so punch-ins stay sharp.
  • Horizontal 1080p gives a 608x1080 crop, scaled up 1.8x. Say it will look softer.
  • Vertical footage works. Punch-ins over 1.2x will soften it.

Step 1. Script and cuts

  1. Transcribe: python $KIT/scripts/transcribe.py source/clip.mp4 -o work/words.json.
  2. Write the reel as a script (rules section 11). Strongest line first.
  3. Plan cuts: stumbles, repeats and fillers inside a thought go. Real pauses between thoughts stay at 0.25 to 0.4 s.
  4. Check every cut: python $KIT/scripts/silence_cuts.py source/clip.mp4 work/cuts.json.
  5. Cut, then run edge_check.py on the joins.

Step 2. Frame the face

  1. Detect the face on every 5th frame. Smooth the track with a 1 s window.
  2. Place the crop so the eyes sit near y 640 of 1920 at 1.0x.
  3. Write the chin line per shot. Captions start 40 px or more below it.

Step 3. The hook (first 1.5 to 3 s)

  1. Start the crop at about 1.25x. Ease it out to 1.0x over the whole hook line. One slow curve.
  2. The hook line goes on screen as big captions in two lines, on frame 0.
  3. Size: the longest line fills about 80% of the frame width, often 110 to 140 px. Line height 0.95.
  4. Split the line at a natural break. Put the punch word at the end of line 2.
  5. Optional name tag: a small lower third with the speaker's name and title, from the user. 0.3 s in, held for 2 s, then out. Never invented.
  6. The hook text is the speaker's exact first line.

Step 4. Punch-ins

  • Three framings: 1.0x, 1.15x, 1.32x.
  • Change framing at every cut, so no join plays as a jump.
  • Go tight on the problem and the key line. Go back to 1.0x on the turn ("So here is what I do").
  • Never go from 1.0x straight to 1.32x and back within 2 s.
  • Bake the punch-ins into the footage frames if the renderer applies them a frame late.

Step 5. Captions

  • Bold sans, white, 2 lines max, 2 to 5 words per group.
  • Centred, under the chin, inside y 220 to 1470.
  • One hero word per group at most, 1.6 to 2x the caption size, in the brand accent.
  • A short shadow for contrast. No box unless the brand file says so.
  • Each group appears on its first word's sound, with the measured lead.
  • No caption shows during a full-screen scene.
  • Every group shows on screen. Check the render.

Step 6. Scene types

Use a scene only where the words describe something you can show. Keep scenes to about 15% of runtime on organic reels.

SceneGood forBuilt from
Key-word pillone idea, one strong worda pill under the captions, on the footage
List stack"three things", stepspills or cards that stack in, one per spoken item
Numbera figure the speaker saysa large number that counts up on the spoken number
Phone"I texted", "I posted", "I got a message"a real-looking phone body and OS screen with content matching the words
Before and aftera change, a contrasttwo cards, the first struck through as the speaker turns
B-rolla place, an actionlicensed stock or the user's own shots, 1.5 to 3 s, with a soft push
Blur returncoming back to the face after a scenefootage starts blurred at about 30 px and 1.1x, clears in 0.45 s

Start from $KIT/templates/ (see the README there) and the brand pieces in brand/pieces/. Fill them only with the speaker's words.

Pills and cards go on the real footage, under the chin. No cut-out of the speaker over a new background on organic reels.

Step 7. Plan, then build

  1. Write work/plan.md: one row per line with time, words, framing, caption groups, scene (if any).
  2. Show the plan. Wait for an OK.
  3. Build the hook first. Render the first 3 s. Wait for an OK.
  4. Build the rest in one HyperFrames project: work/reel/index.html.
  5. Snap each scene's in and out to a footage cut.
  6. Render: npx hyperframes render work/reel. Output 1080x1920 at the footage frame rate.
  7. Mix: voice at -14 LUFS, optional bed 20 LU under, soft pops only.

Checks for this look

On top of the quality gate in video-edit-rules:

  • Frame 0 shows the hook text and the face.
  • Step through each cut. The framing changes. No 1-frame wrong crop.
  • No caption touches the face at any framing.
  • Count the time in full-screen scenes. Report the share.
  • The last line is complete, with 0.4 to 0.8 s of air.

Report

  • File name, length, first line.
  • Framings used and where.
  • Scenes used, on which lines, and their share of runtime.
  • Every sound effect with time and level.
  • The quality gate result.

Individual skills in this repo

This repo contains 3 individual skills — each has its own dedicated page.

Related Skills