Podcast clips
Read video-edit-rules first. Its rules apply to every step here.
Choose the moments with longform-to-clips. This skill is the look.
The goal: one locked wide shot that feels like a multi-camera show. The test: viewers feel several angles, the text sits in the empty space and never on a face, and every cut sits in silence.
Footage that works
- One wide camera, locked off, both speakers in frame or one speaker with space beside them.
- 4K is best. A 2.5x zoom on 4K still gives about 1536 px across, sharp enough for 1080 output.
- 1080p footage limits zooms to about 1.4x before it looks soft. Tell the user.
- Clean voice. Separate mic tracks help with speaker labels.
Step 1. Clip script
- Get the moment and its script from
longform-to-clips: exact words, order, source timestamps. - Mark who speaks each line.
- Mark 2 to 4 key words per clip for the accent. Pick words that carry the idea: a number, a contrast, a result.
- Mark 1 or 2 lines for a full graphic scene. Pick lines that describe something you can show: a list, a before and after, a figure that was said.
Step 2. Map the wide shot
- Grab one sharp frame. Detect each face (OpenCV DNN or MediaPipe). Write each face box to
work/faces.json. - Find the empty space beside each speaker: the largest rectangle with no face, no body and little texture (low Laplacian variance). This is where text goes.
- Plan the framings per speaker in
work/framings.json:
| Framing | Zoom on 4K | Use |
|---|---|---|
| Wide | 1.0x | first line, turns between speakers |
| Medium | 1.5x | most lines |
| Close | 2.0 to 2.5x | key lines and the payoff |
For 9:16 output, the crop is a vertical window into the wide shot. Put the speaker's eyes near y 700 of 1920. Text goes below the chin or in the space beside the head, inside the safe zone.
Step 3. The camera move
Every cut lands on a new framing. Then the camera settles.
- At the cut, start 6 to 10% tighter than the target framing.
- Ease out to the target over 0.6 to 0.9 s. Use one curve (for example cubic-out). No stop and go.
- Hold still until the next cut. A very slow drift (about 1% over 3 s) is fine.
- Never cut between two framings closer than 1.25x apart. It reads as a jump.
- Bake the crop into the footage frames with ffmpeg (
cropplusscale, with expressions per frame) or OpenCV. Do not rely on a page transform for the camera. - Map source frames evenly to the output frame rate.
Step 4. Text in the empty space
Text is pinned to the picture, not to the screen.
- Place each text block in the empty space found in step 2, in wide-shot coordinates.
- Each frame, run the block through the same zoom and offset as the footage. The text grows and moves with the camera.
- Two sizes. A big condensed word (for example 180 to 260 px at 1080 tall) for the key word. Small tight text (40 to 56 px, letter spacing -0.02em) for the rest of the line.
- The text block never crosses a face box, even at the tightest zoom. Check every frame.
- Bring text in on the word, with a short rise (0.3 s) and no bounce. Take it out on the cut.
Step 5. Accent on key words
Two forms. Pick one per clip and keep it.
- Pill: the key word sits in a solid accent pill. Pill height 1.3x the cap height, radius half the height, padding 0.4em each side.
- Sliding box: an accent box slides behind the words as they are spoken, in one smooth move from the first word to the last. It never stops between words. Drive its width from the word times, then smooth it with one easing curve over the whole phrase.
Use the brand accent from brand/theme.css. If there is no brand, pick one bright colour and keep it.
Step 6. Full graphic scenes
Between talking parts, cut to a full-screen scene for 1.5 to 3 s on the marked lines.
- Build each scene as HTML from
$KIT/templates/or the brand pieces. - The voice keeps playing under the scene.
- The scene shows what is said: the list items as they are said, the number as it is said.
- Keep scenes under about 15% of the clip on organic posts.
- Use only pictures the user owns, frames from the recording, or licensed stock. Never invent follower counts, names or figures.
Step 7. Flash on cuts
- On each cut between framings or into a scene, add a light flash: a white layer at 35 to 55% opacity, peak on the cut frame, fade out over 4 to 6 frames.
- Pair it with a soft pop at least 20 dB under the voice, or with no sound.
- No whoosh. No riser. No reverse swell.
Step 8. Build and render
- Footage with the baked camera:
work/cam.mp4. - Text, accents and scenes: one HyperFrames project in
work/clip-NN/. - Snap every scene in and out to a footage cut.
- Render:
npx hyperframes render work/clip-NN. Output 1920x1080 for 16:9, 1080x1920 for 9:16. - Show the first 3 seconds first. Wait for an OK. Then render the whole clip.
Checks for this look
On top of the quality gate in video-edit-rules:
- Step through every cut. The new framing starts tighter and eases out once.
- Step through the tightest zoom of each shot. No text touches a face.
- Watch the sliding box at 25% speed. It never pauses between words.
- Check each flash peaks on the cut frame, not one frame after.
- Check that each zoom is sharp. If a zoom is soft, back it off.
Report
- Clip length, source in and out times, first and last line.
- Framings used and their zoom.
- Key words with the accent.
- Scenes and their share of runtime.
- Every sound effect with time and level.
- The quality gate result.