---
name: talking-head-edit
description: >-
  Use when planning, cutting, adding text to, sound-designing or exporting a
  talking-head video (reel, Short, YouTube, site or sales video). A step-by-step
  recipe from idea to export, with exact timings for the text, cuts and sound.
---
# The talking-head edit (olaceo)

This is the recipe I use on my own talking-head videos. It works two ways:
- **No code:** read it top to bottom and do each step in CapCut (free) or any editor.
- **With an AI agent** (Claude, ChatGPT, Cursor, Codex and so on): drop this file in as a skill or paste it into the chat, give it your footage and transcript, and have it run the steps.

Every number below comes from a real edit. Where a value depends on your footage, it says so.

## 1. Idea
- One video, one person you're talking to. Don't mix audiences.
- One promise: "After this, you'll know ___." Pay it off inside the same video.
- Use something you actually did or learned. No made-up results, no scare numbers.

## 2. Script
1. **Hook:** your first words. Nothing before them: no logo, no title card, no music intro. Call out who it's for ("You are a business owner.") or name a pain they already feel.
2. **Body:** one idea per sentence.
3. **Payoff:** the main gain word.
4. **One ask** that matches the button or the caption.
5. **End card:** about 2.6 s, your logo or name on a dark screen.
- Length: 55 to 90 s for a short.
- Before you shoot, pick one big word per line: the word that carries the point.
- Kill list: "hey guys", "what's up", "stop scrolling", "in this video".

## 3. Shoot
- Phone sideways (landscape), chest height, on a tripod, wide enough to crop in later.
- Soft light in front of you, never behind you.
- Clip-on mic on your chest, away from traffic and AC noise.
- 4K at 60 fps if your phone has it.
- Talk to one person. Finish every sentence on a falling note, then breathe. If you stumble, restart the whole sentence. Pause between sentences so the cuts have air.

## 4. Cut
- Cut stutters, restarts, filler words and dead air between sentences.
- Leave a little air on every cut: about 60 ms before the first sound and 120 ms after the last.
- **Never cut a word.** Never cut while your voice is still going up. Keep the breath at the end of each sentence. Keep a pause that means something (the beat before a big line).
- **New sentence, new framing.** Alternate a wider shot with your face on one side and a tighter shot with your face on the other side. Never the same framing twice in a row.
  - 4:5 feed (1080x1350): wide = 1.0x, tight = 1.3x. Face centre at 30 % or 70 % of the width.
  - 16:9 (1920x1080): wide = 1.45x, tight = 1.85x. Face centre at 36 % or 64 % of the width.
  - Keep your head at the same height in every shot.
  - CapCut: Split, Delete, then scale the next clip to about 130 % and slide it to the other side.
- **Pause cut inside a sentence:** a hard 12 % punch-in on the same side. No transition, no sound.
- **Between sentences:** a short push. 0.2 s ease-in sliding about 4 % of the width with a 1.07 scale and a little sideways motion blur, then 0.28 s easing back after the cut, plus a soft whoosh peaking on the cut frame.
- Inside each clip, an optional slow push-in from 100 % to 103 % about your face.
- Never: dissolves, flashes, zoom transitions, shake, bounce, 3D.

## 5. Text
- **Two fonts:** a small serif for the setup words and one big bold word for the point.
  - Free pair: TeX Gyre Termes (setup) and TeX Gyre Heros Bold (big word), GUST font licence.
  - Setup words: lowercase, about 46 to 48 px on a 1080-wide frame, letter spacing -0.02em.
  - Big word: as said, with its period ("business owner."), letter spacing -0.05em, as large as fits beside your head (up to about 15 % of the frame height in 4:5).
- **Placement:** beside your head on the side away from you, about 6 % in from the edge. Never on your face, never behind you. Check every shot.
- **Timing:** each line starts on the exact frame you say its first word (about 0.04 s early is fine).
- **Entrances:**
  - Setup words: blur-in over 0.4 s (blur 12 px to 0 with a small rise).
  - Big word, rotate these, never the same twice in a row:
    - Flicker: three quick flashes in the first 0.2 s, then solid.
    - Stack: word by word, 0.05 s apart, each rising in over 0.3 s.
    - Box flash: a white box flashes behind the word twice in about 0.13 s.
    - Focus pull: scale 1.35 to 1 with blur 22 px to 0 over 0.22 s.
    - Glitch (key phrase only): a quick red/cyan split that settles in 0.3 s.
- **Exits:** fade out over 0.3 s with a small upward drift.
- White text with a soft shadow. A soft dark gradient behind the text only when the background is bright.
- No full word-for-word subtitles burned in. Use the platform's caption track if you need captions.

## 6. Color
- White by default.
- Red (#E5484D) for a loss, green (#30A46C) for a win. About one colored word every 15 seconds, at most. Never two in a row, never on the call to action.
- The word arrives already colored and stays that color. No color pulse.
- Biggest loss line, once per video: fade the background to black and white and darker (about 45 % down) over 0.45 s, text stays bright.
- Grade: a gentle S curve and a touch more saturation (about +4 %). No filter packs.

## 7. Sound
- **Voice first:** light noise reduction, high-pass at 80 Hz, a small cut around 300 Hz, a small lift around 4 kHz, a de-esser, gentle compression.
- **Room tone:** one quiet, continuous bed of your own room noise under the whole video so cuts never drop to dead silence.
- **Music:** way under your voice (about 14 LU under). It never cuts to silence; dip it instead.
- **Hits:**
  - One soft click when each big word lands, on the word.
  - One soft whoosh on each between-sentence cut, peaking on the cut frame.
  - One low boom per video, on the payoff word.
  - Levels: clicks about 17 dB under your voice peak, whooshes 12 to 16 dB under, the boom about 9 dB under.
  - Don't repeat the same file within 4 hits. 3 to 8 hits per 10 seconds, never more than 8.
  - No beeps, no cartoon sounds, no harsh digital swooshes.
  - Free sounds and music: Pixabay (check each file's licence).
- **Master:** -14 LUFS, true peak below -1 dBTP.

## 8. Export and check
- 60 fps (exactly 60, not 59.94). 1080p, or 4K only if you shot 4K.
- 4:5 for the feed, 16:9 for YouTube. 9:16 works too: put the 4:5 frame in the middle and keep text out of the app's buttons at the bottom.
- Watch it once on your phone with sound on, start to finish.
- First frame: no half-loaded text, no black flash.
- Every word reads clean through every cut.
- Caption: one punchy line, then your ask.

## For AI agents
When the user gives you footage:
1. Transcribe with word timestamps.
2. Write a beat sheet: one row per spoken line with the setup words, the big word, its entrance, color (usually none), the framing (wide/tight, left/right) and the sound cue.
3. Build the cut list on the pauses with the padding above. Never cut inside a word.
4. Place the text from the word timestamps. Check every 0.1 s that no text box touches the face box (widen the face box generously for hair or a hat).
5. Mix the sound with the levels and limits above and measure loudness (ffmpeg `ebur128=peak=true`).
6. Show the user a low-res preview and a contact sheet before any full-quality render.

Free to use and share. olaceo.com