Point it to the course site and generate videos at three levels — overview, tracks, and in-depth lessons by module — in INEMA’s premium dark amber style, with local narration and a CTA for INEMA.CLUB.
scene illus · inemavox narration · captions at the bottom
The engine (engine.mjs) already includes the style, animations, and scene types. Day to day, you only write the specs for each video — the rest is the pipeline.
An overview video — the course landing page and tracks (what it is, numbers, the N tracks). ~45–60s, rachel voice.
One video per track showing modules and topics; an intro for each module is optional. Trailer/roadmap level. ~30–90s, rachel voice.
An in-depth lesson per module (covers the ~6 topics), richly illustrated with SVGs. ~3–4 min, bella voice.
⚠️ The in-depth lesson is per module, not per topic: 6 tracks × 6 modules = 36 deep-dive videos, not 216.
The narration is generated first; the build.mjs reads the duration real for each WAV file and syncs animation and captions. To change the voice, you only change the TTS step — everything else stays identical.
Cloned voice narration (inemavox / Chatterbox-VC) running on 127.0.0.1. Costs nothing and pronounces English terms well within PT-BR.
The generated HTML is a HyperFrames composition with a paused GSAP timeline — rendered to MP4 in 16:9 or 9:16.
title · topic · lead · bullets · cards · steps · term · compare · illus (SVG) · cta. The last one is always the INEMA.CLUB CTA.
The repo includes the installable skill videos-cursos-inema in skills/videos-cursos-inema/ — it teaches the agent to extract the course, write the specs, and run the pipeline. Copy to ~/.claude/skills/ and invoke with /videos-cursos-inema.
Everything runs locally. The only external service is the voice daemon, which runs on 127.0.0.1.
Node 18+ (native fetch, ESM) and FFmpeg with ffmpeg/ffprobe on the PATH.
# check node -v && ffprobe -version
CLI via npx (version pinned in package.json). Check the environment.
# diagnostics npx hyperframes doctor
inemavox daemon running and local sources generated once.
# voice curl 127.0.0.1:7860/health # sources node fetch-fonts.mjs
Each video is an entry in the object VIDEOS (in specs.mjs for Parts 1–2, specs-deep.mjs for Part 3). The ID always starts with the course slug, e.g., fep-deep-t1-m1.
Copy the skeleton and fill in the scenes. The last scene is always cta.
cp assets/spec-template.mjs specs.mjs # edit the VIDEOS object
Generates the text files for each scene from the field narration.
node write-txt.mjs <id> # -> assets/txt/<id>-sN.txt
Calls the voice daemon and saves the WAVs. Use bella (default) or rachel for Parts 1–2.
node tts-inemavox.mjs <id> bella # -> assets/audio/<id>-sN.wav
Reads the actual WAV durations and writes the index.html. No flag = 16:9; --vertical = 9:16.
node build.mjs <id> # 16:9 node build.mjs <id> --vertical # 9:16 (Shorts/Reels)
Lint must report 0 errors. Check 1 frame per scene before the final render — renders take a long time, so run them in the background.
npx hyperframes lint npx hyperframes render --quality high --output renders/<id>.mp4
The defaults cover the common case; each component can be swapped without changing the rest of the pipeline.