---

name: stream-studio description: End-to-end workflow for a livestream/VOD folder. Organizes the stream into a standard folder layout (source, clips, shorts, longform, transcripts, subtitles, chapters, thumbnails) with linked file names, transcribes it, finds each feature/topic and cuts one clip per feature, writes chapter timings (X/YouTube paste-ready), builds summary/highlight reels, and makes animated 9:16 YouTube Shorts (split-screen with the creator's camera, word-synced captions, frame-0 hook, real footage mixed in) with YouTube titles and descriptions, every version reproducible. Use for /process-stream and /make-shorts, or whenever the user wants clips, a transcript, chapters, subtitles, a summary video, shorts/reels or YouTube copy from a stream, keynote, webinar, podcast or recording.

Stream Studio

One skill for the whole pipeline: raw stream → transcript + chapters + clips → summary reel / shorts → YouTube copy. All files follow one folder layout and naming scheme so a clip, its transcript, subtitles, chapters and thumbnail always link together, and every video is rebuildable from its version folder.

SKILL=<this skill's base directory>
$SKILL/scripts/stream.py     doctor · init · transcribe · clips · reel · short-new · short-build · migrate
$SKILL/scripts/short.py      shorts engine front end: doctor · brand · transcribe · frame · cut · scaffold · balance
$SKILL/engine/  assets/      motion-reel renderer, kit.js, fonts
$SKILL/references/           folders.md · process-stream.md · make-shorts.md · style.md · kit-api.md · youtube.md

Which workflow?

the user says	do	read
/process-stream video.mp4, "clip my stream", "transcript", "chapters", "subtitles"	process the stream	references/folders.md, then references/process-stream.md
"summary video of all features", recap, highlight reel	summary reel (longform)	references/process-stream.md → Summary reel
/make-shorts <topic or file>, "make a short", "vertical", "reel"	make shorts	references/folders.md, then references/make-shorts.md (+ style.md, kit-api.md, youtube.md)
"chapter timings I can paste in X/YouTube"	chapters	references/process-stream.md → Chapters
"title and description"	YouTube copy	references/youtube.md

A request often chains them ("process this stream and make a short on dots"): do /process-stream first, then shorts from the clips.

Works in Claude Code and Codex

• The skill is the same in both. /process-stream and /make-shorts are Claude Code commands only; in Codex ask in plain words or mention the skill ($stream-studio process my stream video.mp4, $stream-studio make a 30s short about <topic>).
• Tool names in these docs are generic: "run in the background" = Claude Code's Bash run_in_background, or in any agent nohup <cmd> > .work/job.log 2>&1 & and poll the log; "ask the user" = a question tool if you have one, otherwise a short numbered list in chat; "look at a frame" = open the PNG with your image-viewing tool.
• Python scripts locate themselves ($SKILL = the folder holding this file), so the skill works from any install location.

Always

1. Folder layout first. The user's working directory is usually the stream's root folder. Check for stream.json. If it's missing, run stream.py init (new stream) or stream.py migrate (old layout, dry run first). Never write outputs loose in the root.
2. Name by convention. <slug>--NN-<feature>.mp4, <slug>--short-<topic>--vN/. Details in references/folders.md.
3. Version, don't overwrite. A changed short or reel is a new --vN folder holding its inputs (short.json, page.html, reel.json) so it can be regenerated. Keep old versions.
4. Run stream.py doctor before the first heavy step; it prints the fix for anything missing.
5. Long jobs in the background (transcription, clip cutting, rendering). Don't sleep-poll; do independent work or wait for the completion notice.
6. Ask only what you can't work out, all in one round: brand/handle (shorts), which topics/features, camera setup. Look in the folder first (existing brand.json, previous versions, frames).
7. Be honest about limits. You read transcripts, not audio: you can't hear tone, music or audio levels, only check levels with short.py balance. Whisper-small mishears names. Chapter starts snap to transcript segments (±seconds). Webcams that are small get soft when upscaled. Mock UI and animated counters are illustrative: say so. Don't invent claims, stats, handles or links; quote on-screen facts only.
8. Verify before reporting: look at frames (contact sheet) of anything you render, frame 0 of every short, and the duration. Report paths, durations and caveats; don't paste whole documents into chat.
9. Publishing steps are the user's: don't post to X/YouTube. Give paste-ready text.