# Reference-led visual assets

Use this only when the task needs raster artwork. First inspect the reference and existing assets. Specify what must match: subject identity, pose, gaze, crop, relative scale, palette, light, texture, silhouette and placement. Distinguish a style reference from an exact asset/edit target. Reuse or edit an authorized supplied asset when its exact identity matters; fresh generation cannot guarantee a perfect reconstruction.

Plan the asset for its real page slot: aspect ratio, desktop/mobile crop, quiet space for live text, overlap boundaries and file size. For a layered hero, generate a background plate and a transparent foreground subject only when independent positioning is useful. Keep navigation, type, buttons, charts and simple geometric decoration in HTML/CSS/SVG. Do not bake the working interface into an image.

Use the environment's available generation/editing tool. Verify an available configured image provider/tool before claiming generation is possible, regardless of which agent hosts the skill. Preserve an explicitly requested model; do not silently substitute a provider/model or assume a ChatGPT login supplies API image access. Keep credentials outside project code. Never claim a backend model ID the tool did not expose.

Give each reference image and output a clear role. Request true transparency for cutouts, preserve alpha, and inspect edges against both light and dark surfaces. Keep the same chosen subject asset across related sections; do not generate inconsistent copies. Inspect the output and its actual page composite, including mobile crop and foreground legibility. When a mismatch remains, make a targeted edit with the original asset as input; do not promise exact fidelity or keep regenerating indefinitely. Use the task's existing time/cost budget and disclose unresolved differences.

Save selected assets in the project and record prompts, reference roles, output paths and actual available model/tool metadata. Keep originals and create sibling variants for edits. Reuse on subsequent requests. Generation runs only when the task calls for new or changed artwork, never automatically on every prompt or every quality check.