AI Podcast to Video · skill pública Open Skill
AI Podcast to Video
Audio in. A finished video out — captioned, visualised, chaptered, and cut into the clips people actually watch. This is not an audiogram. It transcribes your recording, finds where the subject changes, generates visuals that follow what is actually being said, and delivers both a full episode and three short clips. WHAT IT DOES • Transcribes the audio and READS it — speaker turns, topics, filler density, and the three best standalone moments • Trims before visualising: removes the throat-clearing intro, dead air over 1.2s, and the worst 10% of filler • Plans one visual per topic, changing every 8–15 seconds — never per sentence, never a frozen frame • Generates stills that relate to the words, with the lower third left clear for captions • Uses motion only where it earns its place: the hook, the reveal, the outro • Burns in captions synced to the transcription, 2–6 words per card, inside the safe area • Places an underscore well under the voice — or leaves the voice alone • Delivers the full episode AND three clips, because clips are what get watched WHAT MAKES IT GOOD • It cuts the audio first. A tighter recording beats any amount of visual effort • At least 30% of visuals are objects, places or diagrams — not people, not abstract gradients • Clips open on the strongest sentence — no intro, no logo, no first two seconds to swipe past • Captions are treated as the reason people watch, because most viewers are on mute GOOD FOR Podcast episodes on YouTube, interviews, voice memos, faceless videos from your own recording, audiograms, and repurposing an episode into social clips. NOT FOR Generating a new script from scratch, music videos, or republishing someone else's audio. Only process audio you own. Upload the recording and say what you want — it will trim, transcribe, visualise, caption and cut it.
Explainer
4 0 1
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4267104/2026-10-20/d7np8exx.png
--- allowed-tools: - generate_image - generate_video - generate_tts - generate_music - transcribe_audio - recognize_video - recognize_image - split_image - mix_clips - download_video - create_document - update_document - memory_whoami - memory_query - memory_store --- # AI Podcast to Video Audio in, finished video out. This skill transcribes the recording, finds the structure, generates visuals that follow what is actually being said, adds captions, and cuts both a full episode and the short clips people actually watch. ## When to use - The user has an **audio file or a link** and wants video. - The user wants a **podcast episode** on YouTube, or **clips** for Shorts / Reels / TikTok. - The user wants a **faceless video** from their own recording. - The user wants an **audiogram** with captions and a visual bed. - The user wants an **interview or voice memo** turned into something watchable. Do **not** use this for: generating a new script from scratch (that is a writing task), a music video, or re-uploading someone else's audio — see Notes. ## Steps ### 0. Intake Call `write_todos`: intake → transcribe → structure → visual plan → stills → shots → captions → cut → deliver. Ask at most **three** questions: 1. **Where is the audio?** — upload, or a link (resolve with `fetch_video_info` / `download_video`). 2. **What is the deliverable?** — full episode, 3 short clips, or both. 3. **What is the visual register?** — talking-head with graphics, faceless b-roll, or animated illustration. If the user only gives a file and says "make it a video", choose the rest and say so. ### 1. Transcribe — everything depends on this ``` transcribe_audio(audio_url=<url>) ``` Then **read it before you plan anything**. You need: - **Speaker turns** — who is talking, and how long each stretch is. - **Topics** — where the subject actually changes. - **The three best 20-second moments** — the ones that work standalone. - **Filler density** — where the "um"s cluster; those are the cuts to make. Save it: `create_document(title='<episode> — transcript', content=...)`. ### 2. Structure — cut the audio before you visualise it Do not visualise a rambling recording. Trim first: - **Remove the first 15 seconds** — most recordings open with throat-clearing. - **Cut dead air over 1.2 seconds**, except before a reveal. - **Cut the worst 10% of filler.** Keep enough that it still sounds human. - **Identify 3 clip-worthy 20–40 second windows** and mark their in/out timestamps. A tighter audio track makes a better video than any amount of visual effort. ### 3. Visual plan — one visual per topic, not per sentence ``` | Section | Time | Visual type | Subject | |---|---|---|---| | Cold open | 0:00–0:20 | Hook card + one shot | The question | | Topic 1 | 0:20–2:10 | B-roll + captions | ... | | Topic 2 | 2:10–4:30 | Illustration + captions | ... | | ... | | | | | Outro | ... | Card | ... | ``` Rules: - **Change a visual every 8–15 seconds.** Longer and the video feels frozen; shorter and it is noise. - **A visual must relate to the words**, not to a generic "podcast" aesthetic. - **At least 30% of visuals should be objects, places or diagrams** — not people and not abstract gradients. - **There is no need to generate a person** unless the user is on camera. Faceless is faster, cheaper and often better. ### 4. Stills ``` generate_image( model='t2i', prompt='<subject from the transcript>, <register: photographic / illustration / diagram>, <palette>, clean composition with space for a caption bar at the bottom, no text', name='V07 — <topic>', entity_type='storyboard_images', orientation='<landscape 16:9 for YouTube | portrait 9:16 for clips>' ) ``` - **Leave the lower third clear** — captions live there. - One subject per image; this is a backdrop for speech, not a poster. - Keep a **consistent grade** across the episode, or it will feel assembled. ### 5. Shots — motion only where it earns its plac