Paste a reel link. The original video plays with its music turned down, your avatar comments in the corner, the numbers appear with their source, and the end invites viewers to the free courses.
reelcomenta is a command-line tool that takes the link of someone else's reel, TikTok or Short and builds a vertical video of your own commenting on that topic. The original video keeps playing, with its music turned down. Your HeyGen avatar speaks in a corner of the screen, the numbers from your research appear in a panel with their source, and the video ends with a question and a call to the free courses at inema.club. It is for people who produce content and want to ride a video that is already circulating, without recording anything. It needs Python, ffmpeg, a HeyGen account with your avatar, and the cardshorts and explicavideos projects.
It was born from the "robot on stage" video (10/09/2026): a robot dancing with a dancer became the hook to talk about jobs replaced by AI and robots.
The first pass is the whole video; then the best excerpt comes back in slow motion while the numbers come in. The original music stays low and drops even lower when you speak.
Each number appears in a panel with a counter, on the exact word where the avatar speaks of it, and carries its source underneath (report and year).
The avatar comes in through the HeyGen studio, automatically cropped to the face, with a frame and a name, without covering the video.
Everything lives in one folder per edition inside edicoes/. The only text you write is roteiro.yaml.
Frame 0 already full: date dd/mm/yyyy, a big headline and the reel running.
3 to 4 cards with an animated counter, a short label and the source.
"Will you keep waiting or will you get qualified?" — "ESPERAR" (wait) is struck through at the turning point.
A mosaic of the real covers of the portal's courses scrolling by, and the final INEMA.CLUB card.
It runs locally. The only outside piece is the HeyGen studio, used through your subscription in the browser.
Python 3 with Pillow, PyYAML and OpenCV (face crop), ffmpeg and yt-dlp.
pip install pillow pyyaml opencv-python yt-dlp sudo apt install ffmpeg
Local transcription with per-word timing and delivery to Telegram come from cardshorts.
git clone https://github.com/inematds/cardshorts
explicavideos brings the HeyGen studio robot (Playwright with your logged-in profile) and the status reader.
# paths via variables, if they are elsewhere export RC_EXPLICAVIDEOS=~/projetos/explicavideos export RC_PERFIL_HEYGEN=~/.cache/inemaccbot/perfil-heygen
In Claude Code or Codex, the reelcomenta skill does all the steps when you send the link. By hand, it goes like this:
Creates the edition folder, downloads the video, makes a 12-frame sheet, transcribes the audio and leaves the script template.
git clone https://github.com/inematds/reelcomenta && cd reelcomenta python3 reelcomenta.py novo "https://www.instagram.com/reel/DeQ62SgBblE/" --id robo-trabalho
3 to 4 numbers with source and year. The speech is 550 to 750 characters, with years and numbers written out in full. Each card enters on a word of the speech (the anchor).
# edicoes/2026-10-09-robo-trabalho/roteiro.yaml (excerpt) fala: | Olha isso. O robô já divide o palco com a dançarina. ... cartoes: - {ancora: noventa, numero: "{} MILHÕES", conta: 92, rotulo: "de empregos eliminados\naté 2030", fonte: "Fórum Econômico Mundial · 2025"} pergunta: {ancora: pergunto, virada: qualific, riscado: ESPERAR, depois: ["OU SE", "QUALIFICAR"]} cta: {ancora: inema, linha1: "+{cursos_100} CURSOS", linha2: GRATUITOS}
Draws the screens with estimated timing over frames of the reel, before spending the avatar queue.
python3 reelcomenta.py previa edicoes/2026-10-09-robo-trabalho # → previa.jpg
Only runs with the approval file. Sends the speech to the HeyGen studio, checks the status and downloads the MP4 (30 min to 1 h in the queue).
echo "aprovado por Nei em 09/10/2026" > edicoes/2026-10-09-robo-trabalho/avatar/APROVADO_HEYGEN python3 reelcomenta.py avatar edicoes/2026-10-09-robo-trabalho
Assembly stops if the speech heard differs from the script. The check makes a 10-frame sheet and measures the volume.
python3 reelcomenta.py montar edicoes/2026-10-09-robo-trabalho # → video.mp4 python3 reelcomenta.py conferir edicoes/2026-10-09-robo-trabalho # → conferencia.jpg python3 reelcomenta.py enviar edicoes/2026-10-09-robo-trabalho "legenda"
A robot dancing on a stage became the hook to talk about jobs replaced by AI and robots. 37 seconds:
| On screen | Source |
|---|---|
| 92 million jobs eliminated by 2030 | World Economic Forum, Future of Jobs 2025 |
| 41% of companies will cut jobs because of AI | World Economic Forum, Future of Jobs 2025 |
| 1 in 5 companies will use AI to cut half of middle management | Gartner, 2026 forecast |
| 542 thousand robots installed in factories in 2024 | IFR, World Robotics 2025 |
| +300 free courses | inema.club catalog |
The first edition validated the format. Next steps:
