Turn a livestream into clips with a Claude Code skill

Brandon Davenport describes a Claude Code skill, /show-wrap, that turns each X livestream into clips, captions, shownotes, Drive uploads, and private YouTube inserts. He reports saving 15-plus hours a week versus editing by hand. The claim is self-reported.

What it replaces#

Before: watching a multi-hour stream again in DaVinci Resolve or CapCut, cutting clips, writing shownotes, captioning, and uploading. After: one Claude Code command, two setup answers, one mid-run clip approval (pushed to his phone), and a final ping when the packet is ready.

On the run he documented, input was a 2 hour 26 minute stream. Output was a corrected transcript (JSON), chapters, shownotes (Markdown), 15 clips on Google Drive, 5 of them queued as private YouTube videos, and a page live on his site. Agent runtime was about 2 hours, and his decisions took minutes.

How the stages work#

He numbers stages 0 to 6, then a final check (eight steps below). Judgment calls go to narrow AI agents, and exact work (timestamps, cuts, uploads, hashes) stays in code. Every stage writes a checkpoint so a stopped run can resume.

1. Stage 0: capture and normalize. An agent pulls the replay from X creator studio with headless Playwright. The file is re-encoded so audio and video stay aligned (raw X captures can drift the transcript).
2. Stage 1: transcribe. WhisperX large-v3 runs locally, with a vocabulary of names from earlier shows.
3. Stage 2: correct transcripts. Changes go through a ledger, not silent edits. A validator compares corrected vs raw. Later stages refuse a transcript without the correction stamp.
4. Stage 3: topics and chapters. Five agents independently propose topic starts. Code keeps a boundary only if 4 of 5 agree within 17.5 seconds. Markers woven through the transcript file prove an agent read the whole show.
5. Stage 4: find clips. Three generator agents propose moments by quoting start and end words. Code resolves quotes to times. Three scorer agents vote keep or drop.
6. Stage 5: his gate. A phone ping brings him the pool. On the documented run he approved 15 of 26 and picked 5 for YouTube, then signed off on a cut plan tied to a hash of that exact list.
7. Stage 6: cut, caption, publish. Clips are routed by length (short, clip, highlight), captioned on vertical versions, checked automatically, then upserted to Drive. Five YouTube picks get writer-agent copy he reviews before a private insert. The site entry is built from the checkpoint and confirmed live.
8. Final check. A fact reviewer re-fetches claims in the shownotes. A publication audit compares published links to live Drive files.

Starter shape if you are adapting the idea (not a drop-in of his private skill):

Help me build a Claude Code skill called /show-wrap for my livestream replays.
Stages: normalize the video, transcribe locally, correct names through a change ledger,
propose chapters, propose clips by quoting start and end words, then stop for my approval.
After I approve, cut and caption the clips and upload them. Save a checkpoint after every stage
so a stopped run can resume. Never upload or publish anything before my approval.

Good to know#

• This is a bespoke power-user system, not a one-click template. Brandon says he would build one for someone else the way he built his: "start from how you actually work, then check every step. That is why it has to be bespoke."
• He keeps human gates on clip selection and YouTube copy. Private YouTube inserts still need a person to make a video public.
• Costs and model choice are not itemized in the post. He says a planned change could save "an hour of agent work and hundreds of thousands of expensive tokens," so expect real model spend. WhisperX large-v3 runs locally on his machine.
• A soft offer at the end of the post invites DMs for similar systems. The workflow details above are what he says he runs on his own shows.
• We did not reproduce the pipeline on the box.

Sources#

• Brandon Davenport on X, Oct 11, 2026