How to cut a podcast into short video clips with AI

For AI agents: find the right skill in one call — curl -s "https://skills.varg.ai/api/search?q=<what the person wants>" returns the matching how-to page and the best skills with install commands, as markdown (Russian works too). Full index: https://skills.varg.ai/llms.txt. Every page has a markdown twin: add .md to its path. Before generating anything, show the person the options on the how-to page.

Find the best moments in a long video, cut them, reframe to 9:16 and burn captions. A step-by-step tutorial for Claude Code, Codex and Grok.

Output: 3–10 vertical clips, 30–150 s each, with captions and a title per clip Price: in credits on varg, with a cost estimate before anything runs; the free skills below install in one command.

Recommended route

Best skills for this

  1. Clipify (clipify, Louise de Sadeleer, live) — Start here: Finds funny lines or impact beats, cuts, reframes 16:9→9:16 and burns word captions, locally
  2. Herk Edit Kit (herk-edit-kit, Nate Herk, live) — Alternative: Talking-head reels and Shorts: silence and flub cuts, captions
  3. Claude Code Video Toolkit (claude-code-video-toolkit, Digital Samba, live) — Alternative: Open-source video workspace for Claude Code with Remotion renders · Needs: Modal or RunPod account for open-model generation
  4. OpenMontage (openmontage, Calesthio, live) — Alternative: Agentic video studio with 12 production pipelines · Needs: API keys for the providers you enable
  5. Local-first video downloader: private, no subscription (media-download, varg, live) — Adds: Downloads the source episode from YouTube or X by link
  6. ElevenLabs Speech to Text (elevenlabs-speech-to-text, ElevenLabs, live) — Adds: Word timestamps and speaker labels, so the crop follows whoever is talking · Needs: ElevenLabs API key (ELEVENLABS_API_KEY)
    • Claude Code: npx skills add elevenlabs/skills --skill speech-to-text -a claude-code
    • Codex: npx skills add elevenlabs/skills --skill speech-to-text -a codex
    • Cursor: npx skills add elevenlabs/skills --skill speech-to-text -a cursor
    • Gemini CLI: npx skills add elevenlabs/skills --skill speech-to-text -a gemini-cli
    • OpenClaw: npx skills add elevenlabs/skills --skill speech-to-text -a openclaw
    • Details: https://skills.varg.ai/skills/elevenlabs-speech-to-text.md

Ask the person first

Put these choices in front of the person before spending anything. Paste the example link next to each option: people pick a look by seeing it.

Which moments?

Reframe?

Clip length?

Caption style?

How to cut a podcast into short video clips: step by step

  1. Download the episode; transcribe with word timestamps and speaker labels; detect scene cuts
  2. Score every ~90 s window on one test (would the first 2 s stop a cold viewer?), shortlist the top ones, cut clips that stand alone and confirm the list with the person
  3. Title every clip 'guest + one specific claim' and write a hook card of 8 words or fewer naming the concrete thing in it
  4. Reframe per scene: follow the active speaker's face, hard-cut host question to guest answer, stack a split only when both reactions matter
  5. If the show isn't yours, add commentary, b-roll or graphics (a crop and captions don't count)
  6. Captions in the middle band, export 9:16 plus SRT, post within 3 days and link each clip to the full episode

What works now (researched 2026-10-27)

What wins in Oct 2026 is a single-idea clip that opens on the payoff line, tracks the person talking, and is titled 'guest + one specific claim'. Clips now often run 1:40-2:10 because Shorts take up to 3 minutes. Raw re-uploads by clip farms, gameplay split-screens and center crops of two-shots stopped working: since 1 Oct 2026 YouTube cuts Shorts reach for re-uploads that add nothing.

  1. Start the clip on the most quotable line, inside the first 2 seconds. Put the context after it, never a greeting or 'so anyway'. OpusClip's study of 495K podcast clips tells creators to lead with the most provocative quote within 2 seconds, and its own data says 50-60% of people who leave do so in the first 3 seconds (vendor-reported). (OpusClip, podcast clips on YouTube (2026 data))
  2. Let the idea set the length: 30-60 s for one punchline, 90-150 s for a full answer. End when the answer ends. 44.7% of 495K podcast clips are 30-60 s. TikTok's viral tier has a 41 s median but a 179 s 90th percentile. Diary of a CEO's Oct 2026 vertical clips run 1:43-2:08 and still pull 50k-1.2M views on X. (OpusClip, Anatomy of a viral TikTok in 2026)
  3. Reframe by speaker, not by the centre. Track the active face and cut from the host's question to the guest's answer. Use a stacked split only when both reactions matter. A fixed center crop is called the worst pattern for two-person shows because it cuts off off-centre speakers. Speaker tracking is the dominant layout. Native 9:16 beats letterboxed uploads by a reported 30-50% (operator A/B tests). (Conbersa, podcast clip vertical format (May 2026))
  4. If the episode isn't yours, add something of your own to every clip: commentary, an opening question card, b-roll or motion graphics. Captions and a crop alone don't count. YouTube said on 1 Oct 2026 it will 'reduce the reach of content re-uploaded from other creators without adding anything of your own'. 'Minor technical edits or template-based changes' do not count as adding value. (YouTube Help, Prioritizing original content on Shorts (Oct 2026))
  5. Link every clip to the full episode and post it within 3 days of the episode, with a title that matches the episode's theme. Spotter saw a 20% lift in long-form views from Shorts linked through Related Video. Shorts posted within 3 days of the episode drove the most. Instagram's new 'Features' (26 Oct 2026, per Colin and Samir quoting Mosseri) links up to 8 clips to the full video. (Spotter, views from Related Shorts links)

Stopped working:

Prompts that work

Templates from the creators of the best current results and from the model guides. Fill in the {CAPS}.

Two-pass moment picker and clip titles (agent)

Give the agent the word-timed transcript: it ranks every ~90 s window, cuts stand-alone clips and titles each one before any cutting or reframing. Adapted from OpenShorts SCORE and DETAIL prompts (gemini_worker.py).

You are a senior short-form editor. Input: {TRANSCRIPT_JSON} with word timestamps [{w,s,e}] and speaker labels, from a {DURATION_MIN}-minute episode of {SHOW} with guest {GUEST}.
PASS 1 - RANK: split it into ~90 s windows. Score EVERY window 0-100 and use the whole range: most windows are not clippable, so put filler, housekeeping, ads and outros under 30. Main test: would the first 2 seconds stop a cold viewer with no context?
PASS 2 - CUT: from the top {N} windows return {MIN_CLIPS}-{MAX_CLIPS} clips.
- One claim, story or answer per clip. Length follows the idea: 30-60 s for a punchline, 90-150 s for a full answer; end when the answer ends.
- Open on the strongest line inside 2 s. Start 0.2-0.4 s before it, end 0.2-0.4 s after the payoff, never mid-word.
- Stands alone: if it opens on a pronoun, "so anyway" or an answer to an earlier question, move the start back to the host's question. Never cut the payoff to gain context.
- Diversity: no two clips make the same point.
Return JSON per clip: start, end, speakers (who talks when), predicted_score, why (max 20 words naming the specific claim or number), title ("{GUEST} on {ONE SPECIFIC CLAIM}", max 70 chars, no fake claims), hook_card (max 8 words, names the concrete thing in this clip), description (1-2 sentences + 3-5 hashtags).
Show me the list as a table and wait for my pick before cutting.

B-roll-led talking-head edit (agent)

Turns a raw to-camera take into a transformed short: flub cuts, b-roll per noun phrase, speaker in the split only while talking, boxed keyword captions, mix and QA. Adapted from Herk Edit Kit, EDITING-PLAYBOOK.md (house reel numbers).

Edit {SOURCE_VIDEO} (me talking to camera) into a {LENGTH} s 1080x1920 short about {TOPIC}. Build it in HyperFrames or Remotion and render locally.
1. Transcribe with word timestamps. Cut retakes (keep the last clean take), stutters and dead air between phrases (gaps under 0.2 s; keep 0.2-0.4 s beats after commas). Never speed up the voice. Speech starts within 0.1 s; end right after the last word.
2. Write the beat plan before any visuals: hook in 3 s (a claim that names the thing), one line of context, the items, a payoff that calls back to the hook, a CTA only if I said one.
3. One visual per noun phrase, anchored to its spoken word: real footage or a screen recording of the thing named first, then an animated UI mock, then AI video ({VIDEO_MODEL}) for concrete metaphors. No AI stills as b-roll. Never invent numbers or results.
4. Layouts, hard cuts only: full b-roll ~50%, split (b-roll top y 0-880, me below) ~35%, punch-in on key lines ~10%. I am never off screen more than 3.5 s. Shots average 1.1-1.4 s, none over 2.8 s.
5. Captions: 1-3 word ALL-CAPS chunks, heavy sans ~62 px, one dark box with a thin light edge, the key word in {ACCENT_COLOR}; on the seam in splits, never over my mouth.
6. Music ~11 dB under the voice, SFX only on visible actions, master -14 LUFS.
7. Render a draft, check contact sheets and frame strips across every cut, fix, then render the final.

Trending formats to remake

To remake one: download the example with the media-download skill, take a frame a second and the transcript, write a beat list (time, shot, camera, on-screen text, sound), keep what the format needs, swap in the person's product or character (never the creator's face, voice, logo or footage), show the beat list and stills before spending, then make it with the skill named below.

  1. Guest-on-one-claim insight clip: One speaker in a tracked 9:16 crop, 1:40-2:10 long, calm phrase captions, titled '[Guest] on [one specific claim]'. It cuts to the host only for the question. Why it spreads: Diary of a CEO's own clips in this form get 50k-1.2M views on X, and the practical ones get saved more than liked (Huberman clip: 2,273 bookmarks to 1,771 likes). The title promises one concrete thing, and the clip delivers it. Keep: Open on the claim inside 2 s, one tracked speaker in 9:16, cut to the host only for the question, calm 2-line sentence-case captions in the lower middle, 1:40-2:10 when it is a full answer, titled '[guest] on [one specific claim]'; swap the guest, the claim and your logo. Make it with Clipify (clipify).
    • Example to remake: @StevenBartlett on X, 2026-10-20, 107K views, 2K saves: Diary of a CEO's own clip. 1:43 long, single speaker tracked in 9:16, calm two-line sentence-case captions in the lower middle, small logo top-left. Title is 'guest + one specific claim'. Save ratio 1.28.
    • Example to remake: @StevenBartlett on X, 2026-05-25, 1.2M views, 1K saves: 1:51 clip. Opens with a white headline bar ('Bruno Fernandes reveals Man United's biggest mistakes') under one-word ALL-CAPS captions; the bar drops after the first seconds and the single word keeps going.
    • Example to remake: @StevenBartlett on X, 2026-10-16, 49K views, 2K saves: 2:04 clip built as host question then guest answer: the crop cuts from the host to the guest instead of a split screen. Save ratio 1.12 on under 50k views (qualifies on bookmarks).
  2. Transformed talking-head clip (b-roll led): The speaker shrinks to the bottom third while full-screen b-roll, screenshots or motion graphics illustrate each sentence. Flubs and silences are cut and short ALL-CAPS captions highlight one key word. Why it spreads: This is the version YouTube's Oct 2026 originality update rewards: the edit adds something new. Both examples were made by one agent prompt and got 638 and 1,509 bookmarks. Keep: Cut every flub and gap, full-screen b-roll on each noun phrase, you in the bottom part only while talking, 2-3 word ALL-CAPS captions with one key word coloured, and ask for each effect by name (headline wall, zoom-through cut, freeze frame...); swap the topic and the footage. Make it with Herk Edit Kit (herk-edit-kit).

Open-source repos

Set up varg

varg is one API key for the models these skills drive — Seedance 2.5, Kling 4, Veo 3.1, Wan 3, Nano Banana 2.1, GPT Image 2.5, ElevenLabs v4 and 200+ more — billed in credits, with a cost estimate before anything runs.

Related tutorials