Rebuild Your Listing Images From Your Own Photos

To rebuild a thin listing gallery from photos you already own, the Listing Optimizer Skill reads the listing through the Selling Partner MCP, seeds new images from its current photos, checks each one against Amazon's image rules grounded by Amazon Agent Atlas, and patches only the images you approve.

The question arrives in Slack every month: "Can the agent use our existing product photos as seeds to make better listing images that still pass Amazon's image rules, and let me approve before anything publishes?" The pain is the long tail: a studio shoot never paid back on a SKU (stock keeping unit) selling twelve units a month. So I handed our agent a thin water-bottle listing: seed from our photos, check before publish, stop at my approval.

ChatGPT, Claude, Perplexity, Microsoft Copilot, and whatever comes next know nothing about your business out of the box, and your competitors use them too. Kuudo gives those same tools your account as it is right now, your judgment running every time, and your call before anything changes. The Selling Partner MCP supplies the live listing and its images, the Listing Optimizer Skill carries the procedure, Atlas supplies Amazon's cited image rules, and approval controls keep the final call yours.

The Listing Optimizer Skill on the Selling Partner MCP seeds new images from the photos already on your listing

The benefit first: you get new images of your real product without a studio. The Skill pulls the listing's current image URLs through the Selling Partner MCP, from Catalog Items or the listing's own image attributes. Those URLs are public, so it fetches the photos and passes them to your image model as references.

The model is yours. The Skill calls Google Gemini or OpenAI with your own key, so your photos go to the provider you chose. Each brief is built from the listing's attributes, such as product type, material, color, and features, plus the rules for the target slot. That is how a size guide gets the listing's measurements instead of invented ones. If you run Amazon Agent Iris, the image MCP, it exposes the same Gemini and OpenAI models as tools.

The water-bottle listing started with one image. One run produced four candidates:

• Main: a sharpened product shot on a pure white background.
• Lifestyle: the bottle in use, in a real setting.
• Size guide: measurements taken from the listing's own attributes.
• Detail shot: a texture close-up that came back too small and failed the size check.

Because the Listing Optimizer Skill runs on one listing or a whole catalog, the long tail gets the flagship's image standard

The benefit here is catalog-wide quality that was never economical before. The Skill's scope is one listing or a whole catalog, so the same seed, check, approve, and patch loop that fixed the water bottle runs on every thin listing. The long tail stayed underbuilt because a studio shoot never paid back on a slow SKU. When the cost per listing falls to a conversation, that math changes.

The standard is Amazon's, not a house style. Amazon's Product image guide recommends at least six additional images and one video. Its guide to adding images says the additional images should show the product in use or in an environment, from different angles, and with different features. Few long-tail listings come close.

When you ask for a batch, the Skill offers a generate-and-review flow instead of pushing images live unreviewed. To pick which listings go first, the Amazon Agent Data layer lands your Selling Partner and Amazon Ads data in a lake you own, and the Amazon Ads MCP reads live ad performance, so an agent can rank thin listings by the sales and ad spend behind them.

The Skill's validator and Atlas-grounded rules check every image, and a person approves what code cannot judge

The check is what makes the speed safe. Before you see a candidate, the Skill runs its validator on it. The validator returns PASS, FAIL, or REVIEW, and it is honest about its limits: anything it cannot measure becomes a REVIEW item for a person. Here is the summary the agent returned for the water bottle:

{
  "asin": "B0EXAMPLE34",
  "sku": "BOTTLE-32OZ",
  "seeded_from": "the listing's one existing main image",
  "candidates": [
    {
      "slot": "main_product_image_locator",
      "kind": "white-background product shot",
      "result": "PASS",
      "checks": {
        "file_format": "pass",
        "color_space": "pass",
        "resolution": "pass",
        "main_white_background": "pass",
        "main_product_fill": "pass (~87% of frame)"
      },
      "review": ["no text, logos, watermarks, badges, or graphics", "accurately represents the actual product", "shows only the product for sale"]
    },
    {
      "slot": "other_product_image_locator_1",
      "kind": "lifestyle scene",
      "result": "PASS",
      "review": ["accurately represents the actual product", "no accessories the buyer won't receive"]
    },
    {
      "slot": "other_product_image_locator_2",
      "kind": "size guide from the listing's attributes",
      "result": "PASS",
      "review": ["printed measurements match the listing's attributes"]
    },
    {
      "slot": "other_product_image_locator_3",
      "kind": "detail / texture shot",
      "result": "FAIL",
      "checks": { "resolution": "880x880, under 1000px minimum" },
      "next": "regenerate larger; do not patch"
    }
  ],
  "awaiting": "your approval of each image before the patch preview"
}

The validator checks what code can measure. That covers file format, color space, and size on every image. On the main image it also checks for a pure-white border and estimates product fill against an 85% target. The detail shot failed on size: the Skill fails anything under 1,000 pixels on the longest side, which is stricter than Amazon's 500-pixel minimum, so every image it patches supports zoom.

The rules behind those checks are Amazon's, retrieved through Atlas rather than recalled from training data:

• Main image background: pure white, RGB 255, 255, 255.
• Main image content: the product at 85% of the image, the entire product in frame, one unit, and no text, logos, borders, color blocks, or watermarks.
• Size and format: 500 to 10,000 pixels on the longest side, 1,000 or more enables zoom, and JPEG, TIFF, PNG, or non-animated GIF files.
• Accuracy: the image shows the product's real scale, quantity, and color.
• Consequence: Amazon's Image suppression page says that if an image does not comply, Amazon may suppress the listing from search until you provide a compliant image.
• AI-generated people: a photorealistic person generated entirely by AI must be tagged contains-synthetic-performer in the image's XMP (Extensible Metadata Platform) dc:subject field before upload.

Category rules change the main image too. Human models are allowed only for adult apparel, and children's and baby undergarments, leotards, and swimwear must be shown without models. The Skill reads Amazon's Product page style guides through Atlas and never invents a rule it cannot find.

What code cannot judge goes to you. Text, logos, focus, and whether the image shows the product the buyer receives are REVIEW items. A model will happily add a feature, a part, or a badge the product does not have, which is why approval is mandatory in this Skill, not advisory.

What happens next: preview the patch, submit on your confirmation, and re-read the listing

After you approve an image, the Skill stages the file at a public URL in your own storage. Amazon's Listings Items API takes an image as a media_location URL that Amazon fetches, from a public HTTP or HTTPS address, a public Amazon S3 bucket, or a private S3 bucket you grant Amazon access to. The Skill then previews a patchListingsItem call on the chosen image attribute with mode: VALIDATION_PREVIEW, which persists nothing.

You confirm, and the Skill submits the same call without the preview mode and reports the submission ID. ACCEPTED means the submission validated, not that the image is live. Amazon says an image can take up to 24 hours to appear. Uploading also does not guarantee display, because Amazon selects and arranges images from multiple selling partners. The Skill offers one later re-read of the listing's status and issues rather than assuming success.

Three paths lead into and out of this loop:

• A FAIL result goes back through generation and the validator, or the Skill tells you why it cannot fix it.
• A thin gallery found in a listing audit starts here when the audit-to-patch workflow flags an image problem, so it gets rebuilt the week it is found.
• A suppression with no image cause belongs to the suppression diagnosis in the next guide, not to another round of images.

The pattern is the one that makes any of this safe: seed from your own truth, check before publish, approve before live. That is not model cleverness. It is the product surface doing its job.

Next: when a listing stays suppressed after its images pass, listing suppression diagnosis reads the exact reason before any fix.