Dictation and prompt improvement

In the bottom row of the generation panel, next to the launch button, there are two text tools: the microphone and the prompt assistant. On a phone they are in the bottom row of the generation panel. Both buttons are available in any mode where the model has a text field.

The microphone records up to a minute of speech. Recording stops with the Stop and transcribe button or on its own after 60 seconds, then the text is transcribed and appended to the end of the prompt after a space, and what you wrote earlier stays. You do not need to choose a language. The same button is in the Agent chat and in Studio, in the Avatars and Products sections.

The assistant is the button with the Write a prompt for the selected model tooltip. It takes the text in the field as the task and writes a full prompt for the selected model from it. Every explicit fact from your text is kept: the character, the action, lines, text in the shot, song lyrics. The assistant does not add a plot, product or event of its own.

The assistant adds what is missing for execution based on the type of result. For an image: composition, environment, light, palette and materials. For video: action over time, framing, camera behavior and sound. For music: style, tempo, instruments, structure and dynamics.

The assistant sees the selected model and its settings: format, duration, resolution, sound, number of outputs. The values of these settings are not carried into the text; the panel controls manage them. Attached references and slot materials also go into the context. If the model requires references like @Image1, a result without them is not inserted.

The model sets the prompt language. For most models the assistant writes in English, and for Nano Banana, Qwen Image 3 and Seedance it writes in the interface language.

The finished prompt goes into the field, and a Restore previous prompt button appears next to it: it returns the text and settings to their state before the assistant. If the result does not pass the check, because it came out empty, longer than the model's limit or lost a required reference, the field stays as it was, and a hint names the reason.

The assistant also works from the project media library. On a generation card that the filter rejected or that came back empty, there is a Repair prompt with AI action: it rewrites the original prompt and opens the panel in that generation's mode. In the finished image viewer, the Extract prompt button builds a standalone prompt from the content and style of the image.

The Prompt improvement field in project settings sets standing conditions for every assistant run in this project, for example no cartoon look, English text, photorealism.

Separately from the assistant, the panel checks the text against the settings. If the prompt says 15 seconds but a different duration is selected, mentions @image2 that is not attached, or describes video while image mode is selected, that place is underlined, and a hint names the mismatch, for example Duration doesn't match the prompt. Clicking the hint fixes the setting. The hint does not block the launch.