amazon-image-studio

# Amazon Image Studio Amazon-specific knowledge for generating listing images with the general `vibe_seller_generate_image` tool. The tool itself documents the generic contract (user-language prompt, references carry a real subject's appearance, roles by position, the user-confirm pause) — this skill adds only what is Amazon: ## 1. Amazon image requirements **MAIN image** (the one search results show — strictest): - Pure white background, exactly RGB (255,255,255) — off-white fails Amazon's automated scan. - Product fills ~85% of the frame, fully visible, not cropped. - NO text, logos, badges, watermarks, borders, props, or accessories not included in the purchase. Product only, as the buyer receives it. - ≥1000px on the longest side; ≥1600px recommended (enables zoom). Square (1:1) displays best. JPEG/PNG. - Category variations exist (e.g. adult apparel is usually shown on a model; shoes as a single shoe at an angle). When in doubt, mirror what the store's own live listings of the same category do. **Secondary images** (gallery slots 2-7+): lifestyle shots, infographics with feature callouts, dimension/scale charts and comparison tables are all allowed and convert well. On-image text is fine HERE (never on the main image) — spell every word exactly in the prompt and proofread the result character by character. ## 2. Collecting reference images - **Supplier page (e.g. 1688)**: extract the original gallery image URLs from the page (full-size, not thumbnails) and pass them as `reference_images` directly — the generator fetches URLs itself. - **Amazon listing** (style reference): open the listing's dp page and take the hi-res image URLs (`m.media-amazon.com/images/I/…`, request the large `_SL1600_` variant). These carry the store's composition, palette and infographic layout. - **Never pass a blank/placeholder image.** Two traps produce them: - listings whose images were never uploaded show a "No image available" placeholder; - lazy-loading pages serve tiny stand-in GIFs (e.g. 60×40px) until the image scrolls into view. Judge what you actually fetched — is it a full-resolution product photo? A placeholder passed "just in case" poisons the generation. - **Style-reference search order** (autonomous, never stall): 1. a live-imaged listing of the same product type on this store; 2. otherwise ANY live-imaged listing on this store (brand style — hero look, palette, chip band — carries across categories; you reference its composition, never its product); 3. if the whole store has no live images, ask the user once (AskUserQuestion): provide a style image (they can drag one into the confirmation popup's reference area) or proceed with supplier photos only — optional; proceed immediately if declined. ## 3. Generate, audit, hand off - One `vibe_seller_generate_image` call per image (`kind`: "main", "infographic", …). Use `nano-banana-pro` for anything with on-image text. - **Every image comes from the model — including "just remove the background".** "Keep the product, only drop the background", "put it on white", "crop it square" are all *generation* jobs: one `vibe_seller_generate_image` call whose `reference_images` is the photo and whose prompt says to keep the product identical and change only the background. Do **not** reach for `rembg`, Pillow/OpenCV compositing, or ImageMagick — a PreToolUse guard blocks them, and `generated_images/` is the tool's directory (each file's recorded prompt + model is what the user's inline card claims). A local cutout also loses what the model gives you for free: relit edges, a clean contact shadow, and no halo on hair/mesh/glass. - After each image is saved, view the returned workspace `path` and compare against the ORIGINAL supplier photo item by item (shape, colour, texture, proportions; infographic wording). Regenerate with a prompt naming the specific fix if anything differs. - **If you cannot actually SEE the image** (the Read comes back as something you can't look at — some models are text-only), say so and let the user judge: the image renders inline in the task. Do NOT substitute a pixel-measurement script for looking, and never act on what such a script measured — corner-RGB sampling and coverage-percent arithmetic answer a different question than "is this the right product image", and "fixing" those numbers locally is how the deliverable stops being the model's. - **Off-white background or an over-filled frame** — the two things a first pass gets wrong on a main image (e.g. RGB(254,254,253) instead of pure white, or the product bleeding to the edges). Fix by regenerating, never by post-processing: call the tool again with the PREVIOUS generated image as the `reference_image` and a prompt naming only the fix — "keep the product pixel-identical; background must be pure white RGB(255,255,255); leave a clear margin on all four sides so the product fills about 85% of the frame and touches no edge". - **Revising on user feedback** — for ANY change the user asks after seeing a generated image (lighter/darker, bigger/smaller, recolour, remove or add an element, change composition, …): call `vibe_seller_generate_image` again and ALWAYS pass the PREVIOUSLY GENERATED image (its `generated_images/…` path) as a `reference_image`, with a prompt describing ONLY the requested change. This EDITS the prior result — keeping what the user was happy with — instead of regenerating from the supplier photo and drifting. If the user also drops a NEW photo for the change (e.g. "this one's good, add my dog in the middle"), pass BOTH the previous generated image AND the new photo as `reference_images` and name each image's role by position in the prompt ("image 1 is the current design to keep; image 2 is the dog to add in the center"). - Putting the images ON a listing is the `amazon-listing` flow (Manage Images / flat file) — this skill only produces the files.