Podcast promo

The person brings an episode (video or audio; a podcast, an interview, a talk or a live session), the show's branding, who the guests are, and how many clips for which platforms. You deliver that many promo clips of 30–60 s: the best moments with a hook up front, the speaker framed for the platform, captions, the guest's name, the show's bug and a card that sends viewers to the episode. Local and nearly free; the paid part is speaker-labelled transcription when several voices must be told apart. Record the picks, the bites' source times and the brand answers where you keep the job's notes.

Steps

1. Transcribe and find the speakers (guide-select-highlights-from-long-footage): route B, diarized (elevenlabs-speech-to-text, mapped to names with python-read-word-timings turns), when the clip is an exchange; route A (whisper-cli-transcribe-with-word-times) for one voice. Watch the video too (python-contact-sheet): who is on camera when, the reactions worth showing.
2. Pick the moments: per clip a hook a cold viewer gets, then whole sentences on one idea, ending on a payoff (guide-select-highlights-from-long-footage, Process). Gate: the person approves the moments as text.
3. Cut the bites: every edge in a natural pause (python-find-speech-pauses, snap), each checked by its own transcript (ffmpeg-cut-audio-window), and each levelled with one fixed gain (ffmpeg-normalize-loudness, Tips and traps).
4. Frame the speaker (ffmpeg-crop-to-vertical): a 9:16 window on whoever speaks, switching on the diarized turns, never mid-word; a webcam bubble or a small inset is cropped inside its true edge, measured on a full-size frame. Audio only: an audiogram, the guests' photo or the show art with the bite's own waveform scrolling past a playhead (hn4a-compositor-audio-waveform).
5. Brand (guide-capture-a-brand): the show's fonts, colours and logo.
6. Graphics, one spec per clip (guide-add-titles-and-motion-graphics):
   • karaoke captions, the keyword on the accent pill (guide-caption-a-video, 9:16);
   • the show's bug top left and the episode or event tag top right ("LIVE · CITY · DATE" in a thin outlined box);
   • the guest's name strap the first time they speak (hn4a-compositor-chyron);
   • stickers on the key phrases: the phrase in heavy caps on an accent slab, tilted a few degrees, slamming in on its word (hn4a-compositor-in-scene-titles);
   • a designed title card between moments: the big question on the brand colour, a small line under it (hn4a-compositor-motion-graphics);
   • an end card with the episode's name and where to listen. Gate: the live spec (hn4a-compositor, Live spec, then render) on the first clip; a yes on the look before building the rest.
7. Sound: an optional bed under the cards (elevenlabs-generate-music, or the show's own theme), ducked about 12 dB under the voices; pre-mix bites and bed into one track and limit it (ffmpeg-normalize-loudness, Tips and traps); loudness to −14 LUFS for social.
8. Render, check, deliver per the core skill: frame count, a 2 fps contact sheet, the first seconds listened to.

Choices this workflow makes

• The moment is the clip: no generated picture, nothing that changes what anyone said.
• Diarized when it's an exchange, so framing and name straps follow the voice.
• Fixed gain per bite, a limiter on the mix: per-clip loudness pumps short bites.
• The bug and the tag on every clip, so a clip seen alone says whose show it is.

Dead ends

• loudnorm on 2 s bites: pumped at each clip's start and end, heard as distortion.
• The compositor mixing bites and bed as separate tracks: the sum peaked at +2.5 dBFS. Pre-mix and limit.
• A bubble cropped from an estimate showed the top of the source's own webcam circle; it was 13 px off. Measure the circle on a full-size frame and crop inside it.
• A bite that sounded clipped: it started mid-breath. Bites start in silence and end in their pause.
• A bubble showing a loading spinner: the source's webcam reloaded after a click; the bubble pops in only once the face is back.

Worked example

A 30 s 16:9 promo for a podcast-planning app, built from its host's 6-minute screen-recorded demo (1920×1058 at 28.28 fps, a 170 px webcam circle bottom left). Nine scenes cut on 2 s bars of a 120 BPM bed; four bites of 1.7–2.9 s of the host, each with its bubble cut so the mouth lines up.

Spend: a few transcription credits; everything else local.