Reveal a prompt one word at a time, predict what each word changes, then compare against the simulated output — guidance made legible, token by token.
Live preview · launch for the interactive version
Add the prompt word by word and watch the simulated image respond to each addition.
Each word logs what it steered — and the gaps the model is still filling on its own.
See which parts of the image you specified and which the model supplied by default.
Predict what the next word will change before you reveal it.
Add a single word; keep the rest of the prompt fixed.
What moved? What did the model fill that you never asked for?
Say what the word guided and what stayed a default.