heditor: talking-head footage in, a styled 9:16 Reel out, built on HyperFrames
A brand-new MIT-licensed agent skill that cuts a 30–70 second argument out of a lecture, live stream or podcast and rebuilds it as a designed vertical short, then checks its own work.
heditor, from Brazil's PycodeBR (Felipe Azambuja), was created on Oct 13 2026 and had 68 GitHub stars when we checked on Oct 15. It was still being pushed to that day. It's a SKILL.md package for Claude Code, Codex, OpenCode, Cursor and Hermes. The README is in Portuguese with an English section, and the style names are Portuguese too.
The job: you hand it a recording of someone talking (a class, a live stream, a podcast, a talking head) and it returns a 1080×1920 Reel/Short/TikTok with motion graphics, synced captions, sound effects and music. It runs locally. It finds the speaker's framing, transcribes with word-level timing, and picks one complete 30–70 second thought without inventing any speech. Then it cuts speech and camera to the frame, shortens pauses, and treats the voice.
What makes it more than a caption tool is the named styles. 'game' turns the screen into a retro RPG with a HUD and dialogue-box captions, 'thriller' is security-camera tape with REC and a heartbeat, 'ide' is a code editor with a webcam bubble and captions typed as comments, and 'podcast' is a talk-show set with giant karaoke captions. There are also ad formats (keynote stage, stories, abandoned cart, order tracking) and 'authority' looks for expert speakers. Creative mode pitches 6–8 format concepts and changes the screen every 1.5–2 seconds. A reference mode researches a look you name and builds an original homage, and the README says that means no logos, brands, characters or protected music.
The craft details worth copying: each video gets its own DESIGN.md (scene and sound script) before composing in HTML + GSAP and rendering with HyperFrames. The mix is voice-first, with music and effects underneath and mastered to −14 LUFS. QA is unusually thorough: lint, vision-reviewed previews, a top margin measured in pixels, loudness and true-peak checks, a transcription of the final audio, and a 40-frame contact sheet.
Setup is a one-line installer (or npx skills add pycodebr/heditor -g). It installs ffmpeg, Node, a Python venv with faster-whisper and MediaPipe, a pinned HyperFrames version, and a Mixkit library of tracks and effects. As always, read an install script before you pipe it to bash. We haven't run it yet.