python make.py new <name> --from-image <path> (lowercase-dash name).
python make.py check <name>, then read out/<name>/check/grid.png. Find the mouth centre and both eye centres in the grid's pixel coordinates; sample the skin colour next to the mouth (PIL getpixel).
Edit mascots/<name>/mascot.py: MOUTH = dict(at=..., width=..., cover=dict(color=<skin>, size=(w, h))) when the picture already has a mouth, EYES = dict(left=..., right=..., size=(rx, ry), lid=<skin>) for blinking. Width is roughly the drawn mouth's width x 1.1.
python make.py check <name> again and read model_sheet.png: the 9 mouths must sit on the face and fully hide the original mouth. Adjust and repeat until they do.
Voice: ask which provider (macOS say is zero setup; ElevenLabs needs ELEVENLABS_API_KEY in .env; Voicebox / own recording = VOICE = {"provider": "file"} + --audio). For a new ElevenLabs voice use design-voice, let the user listen, then save-voice.
python make.py video <name> --text "...", then python make.py sheet <name> <take dir> 0.5 2 4 and read the sheet before reporting.

Existing mascot

python make.py video <name> --text "..." (or --script file.txt, or --audio file.wav --text "same words"). More poses or hand-drawn mouths: see mascots/_template/README.md for file names; trigger words live in TRIGGERS.

Fixes

Mouth too big/small or off-centre: MOUTH.width / MOUTH.at.
Old mouth showing through: grow cover.size or fix cover.color.
Lip sync feels loose: bash scripts/install_rhubarb.sh (the fallback is loudness-only), and pass --text with audio.
Captions wrong words: pip install openai-whisper.

mascot-video

Mascot video

New mascot from a picture