imagegen

Image generation

Use the built-in ImageGen and ImageEdit tools. Provider authentication, model routing, output storage, and secrets are managed by the desktop host; never ask the user to put an API key in this skill or in the prompt.

Decide the request shape

• Treat a brand-new visual as generation and call ImageGen. Its schema intentionally has no image-path argument.
• Treat a request that preserves, combines, or changes an existing visual as an edit and call ImageEdit.
• One distinct prompt equals one tool call.
• Use count only for multiple variations of the same prompt. For different concepts, make separate calls.
• ImageEdit requires referenced_image_paths: populate it with ordered, exact paths to images the user supplied in this conversation — a path surfaced by [Image source: ...], a file the user attached with @, or a path returned by an earlier ImageGen call. Never invent, search for, or substitute another filesystem path, and never read an image off disk yourself to use it as an input; if the user means an image you have no path for, ask them to attach it. The first image is the primary canvas unless the user says otherwise.
• For multi-turn editing, use the latest selected output as the next turn's edit_target. Repeat all identity, layout, text, and unchanged-region constraints on every turn so edits do not drift.
• To edit several images independently, make one call per image. Put multiple images in one call only when the user wants them combined or used together as references. A single call accepts at most three source images.
• Prefer a useful default composition when the user leaves details open. Do not invent branding, logos, or people they did not request.
• Provider and image model selection come from the current desktop session; do not add either to the tool arguments.
• If the provider returns an error, do not retry the image tool automatically. Explain the failure and let the user decide whether to retry or change providers.

Build the prompt

Turn the request into a complete art-direction brief. Preserve all relevant user-specified detail.

• use case and image type
• subject, action, and important attributes
• environment and context
• composition, framing, and camera angle
• lighting and mood
• visual style or medium
• color palette
• exact text, only when text must appear in the image
• constraints and elements to avoid