Source: raw/How_to_Make_Viral_Shorts_With_AI_Claude_Video_Workflow.md (AI Master, YouTube 4aUVjc7ttko, fetched 2026-10-30). Sponsored by Riverside, the only tool shown; the presenter’s verdicts are part of a paid video.
A step-by-step workflow for turning one 20-minute recording into a batch of vertical shorts inside Riverside. It runs in order: Magic Clips as a first pass, transcript editing to tighten the opening, the Riverside MCP so Claude or Codex can cut clips from a written brief, cleanup (filler words, Magic Audio, captions, framing), and the virality score for posting order. The presenter’s conclusion is that the auto-clipper is only a map, and prompt-to-clip is the step that adds judgment.
Key Takeaways
- Magic Clips is the baseline, not the output. Point it at the full recording and it returns captioned, framed vertical clips with no instructions. The presenter says an auto-clipper “picks moments that sound complete” but “has no idea which moment a stranger needs to hear first”. Treat the batch as a map of where the energy sits.
- Three checks for any auto-clip: does it start on a real hook rather than mid-breath; does the point make sense without the previous 10 minutes; does it end on a payoff? Keep the two or three that pass. Use the rest as timestamps worth a second look.
- Hunt in the transcript, not the timeline. Every Riverside recording has an editable transcript; selecting text makes the clip. Clip-worthy moments read like a strong opinion, a surprising number or a short story with a turn.
- The biggest single fix is cutting the warm-up. “Most weak clips aren’t weak because the idea is bad. They’re weak because the first 2 seconds are warm-up.” Delete words until the clip opens on the claim.
- Fix names in the transcript before export. Captions come from the transcript, so a corrected guest name or product term carries into every clip cut from that passage.
- The Riverside MCP connects recordings to Claude or Codex. The assistant reads transcripts and returns a finished vertical clip when you describe the kind of moment you want (for example, “the spot where the guest pushes back on common advice”) instead of a timestamp.
- A good clip request names three things: the kind of moment, the length, and where the clip must start (usually on the claim). One request can ask for several clips, each built on a different angle from the same episode.
- Known limit: the model reads text, and text doesn’t carry tone. Sarcasm can read as sincere and a sharp line can sound flat. Its picks are candidates; you make the final call.
- Filler removal can flatten timing. Run the automatic removal, then check the cuts around the lines that matter: “the beat right before a punchline is technically a pause.” About a minute per clip.
- Audio and captions: Magic Audio evens levels and reduces background noise. Shorts are often watched on phone speakers in noisy places, and “clean audio won’t rescue a weak clip, bad audio will sink a good one.” AI-generated sound effects (keyboard typing, a notification chime) are made on request, not pulled from a library. Keep captions large and out of the bottom of the frame where platform buttons sit. With two speakers, check the vertical crop follows whoever is talking.
- Use the virality score as a tiebreaker. It is Riverside’s AI estimate of short-form performance. Post the higher-scoring clip first when two feel equal. When score and gut disagree sharply, re-listen to both. The score judges clips in isolation and cannot see your channel history: “A score is a prediction and your channel history is the answer key.”
- Schedule from the same chat. Through the MCP, the presenter says the same conversation can schedule clips across platforms and query which past shorts performed best and what they shared (topic, length, opening line).
- The one human step: listen to every finished cut on headphones before scheduling, about 30 seconds per short. “Every tool in this chain reads text or waveforms. None of them can hear whether a line actually lands.”
Try It
- Record or upload one long-form episode to Riverside and run Magic Clips. Score each clip against the hook / stands-alone / payoff checks and keep the survivors.
- Read the transcript for the episode. Fix guest names and product terms first, then select two or three passages that read as an opinion, a number or a story with a turn. Trim each until the first word is the claim.
- Connect the Riverside MCP to Claude or Codex and send a batch brief. A starting shape based on the video’s three elements:
From the episode "<title>", cut 4 vertical clips, each 30-45 seconds: 1. The contrarian take: where the guest disagrees with common advice. 2. The most practical tip: a step a viewer could do today. 3. One sentence that works with zero context. 4. The strongest number or result mentioned. Each clip must start on the claim itself, not the lead-in. - Run filler-word removal, then manually re-check the pauses before punchlines. Run Magic Audio. Check caption size and position and the speaker crop.
- Sort by virality score, spread the batch over one to two weeks with the highest scores first, and listen to each clip once on headphones before it goes into the queue.
- After the first few posts, ask the assistant which shorts performed best and what they had in common, and use that to brief the next recording.
Implementation
Tool/Service: Riverside (recording, transcript editing, Magic Clips, AI tools pass, Magic Audio, virality score) plus the Riverside MCP connected to Claude or Codex. Setup: The video shows the MCP in use but not the connection steps. In Claude, a remote MCP is normally added under Settings → Connectors → Add custom connector. ^[inferred] Cost: Not stated. The sponsor offer is a creators.riverside.com/aimasterclips link with code “master” for one month free (the video’s claim; check current terms). Integration notes:
- Riverside’s virality score plays the same role as Higgsfield’s Virality Predictor: a pre-post prediction to set order, not a verdict.
- The text-reads-but-can’t-hear limit applies to any transcript-driven clipper, including video-use and Higgsfield Katana in Higgsfield MCP.
Open Questions
- The Riverside MCP’s connector URL, tool list, plan requirements and whether scheduling covers every platform Riverside publishes to are not shown.
- Riverside’s plan prices and whether the MCP and virality score are on every tier are not stated.
- What signals the virality score uses is not explained.
- The “structure I’d give the assistant for a first batch” is shown on screen but not read aloud; the brief in Try It is rebuilt from the three elements the presenter names.
Related
- HeyGen Instant Highlights V2 — HeyGen’s equivalent auto-clipper for long-form to short-form
- Higgsfield Virality Predictor — another pre-post scoring tool, with hook and hold-rate sub-scores
- Higgsfield MCP — Katana, an agent-driven clipping editor inside Claude
- B Testing — reading the post-publish numbers that replace the prediction
- claude-watch — a Claude skill for studying why a viral hook works
- video-use — transcript-driven editing in Claude Code
- Connectors — how remote MCP connectors attach to Claude