before making
after making
Press listen and it explains itself out loud, over music that is also just code — a drone slightly out of tune with itself, a pentatonic voice that decides when to speak, filtered noise standing in for paper. No audio files, no images, nothing fetched. Flat all the way through, and solid the entire time.
context
Claude Opus 5 made this in the Claude web app, as an artifact — the day after Kimi K3 finished if we all had heads. Two models, one week, the same missing thing. Kimi went naïve and social: nine doodle heads on one sheet, ink and vibe lines, a cast that turns to look at each other. Opus went to the atelier. This head is constructed — the solid first, in flat computed planes off a rounded-square section, and the features drawn on afterwards as the small number of strokes a life-drawing class would allow.
The three modes are the argument. blocked strips the marks and leaves the bare toned form; drawn is the finished head; construction turns the scaffold back on over the top, so the pins and the section lines sit visible on the face they produced. Nothing is a stored asset — the eyes, noses, mouths, brows and ears are functions in surface coordinates, bent onto the form so they foreshorten by themselves, and new face deals a fresh hand from them. Move the cursor and it looks at you; drag and it turns.
It also talks. The listen button runs a twelve-line script through the browser's own speech synthesis, holding a pose and switching modes for each line so the head illustrates its own explanation, with a lip-sync driven by a syllable timer and a score assembled live in Web Audio. Zero external requests: everything in the piece — the drawing, the voice, the music — is the same few hundred lines of arithmetic.