ngs-amazon-image-studio

NGS Amazon Image Studio

Goal

Based on the product business card submitted by the user, generate one, multiple, or a full set of cross-border e-commerce product images. Only the following six categories are supported:

White background image white-background
Display image showcase
Selling point image selling-point
Hero image hero
Detail image detail
Scene image lifestyle

The full set mode includes 1 image of each of the above six categories by default. When the user has not explicitly requested a full set, do not expand it into a full set on your own.

This Skill is not responsible for A+ pages, does not upload to Amazon or other e-commerce backends, and does not generate independent size charts, ingredient charts, model images, HOW TO USE images, or usage comparison images. Size, ingredients, people, and usage steps are handled as image content only when the six image categories selected by the user truly require them, not as independent image entry points.

Mandatory execution order

Strictly follow the order below; do not throw out the three questions of business card, image selection, and output path all at once from the beginning:

产品业务卡
  -> 用户选择生什么图并确认尺寸
  -> 用户指定或确认输出位置
  -> 建立产品事实锁与主控参考图
  -> 建立统一视觉概念与创意母题
  -> 输出逐图规划和文件名
  -> 用户明确确认
  -> 按规划逐张生图
  -> 质检、保存、回报文件路径

Step 1: First collect the product business card

On first trigger, first read and use references/business-card.md。

If the user has already provided some fields, organize the existing information and only ask follow-up questions for the missing required items, so the user does not have to fill them in repeatedly.
Product image, product name, and product category are required. The category is used to determine the factual boundaries and the output directory naming.
When the user provides a competitor link, ASIN, or reference image, follow references/competitor-benchmark.md to perform controlled breakdown, produce a competitor insight card, and incorporate it into the unified visual concept.
When the user provides a multidimensional table or document containing prompts, read the prompt field for the corresponding type of this product in that record as the prompt source for Step 8.
When the user provides no reference at all, do not decide arbitrarily: when asking about image type and size in Step 2, you must also ask whether to search online for references. After the user agrees, you must search online for excellent examples in the same category (Amazon top listings image sets, A+ modules, industry cases), observe one by one how display images, selling point images, hero images, detail images, and scene images are respectively made, extract insights, and then create the unified visual concept; if the online results are insufficient, state this directly and do not fabricate examples. When the user explicitly says not to look for references, complete the concept design based on category common sense and independent creativity, and clearly explain the source of the concept in the planning description.
Insights from competitors and online examples only affect display images, selling point images, hero images, detail images, and scene images; the white background main image does not reference competitors or examples and is based only on product facts and platform compliance.
Other fields may be left blank. First make targeted additions or deletions to the business card based on the product category, product structure, and current round's goal, then only ask follow-up questions about fields that truly affect image generation.
Do not mechanically apply the same complete business card to all categories; skincare products, clothing, electronic products, household goods, pet supplies, etc. should use different supplementary fields.
Only when the image type the user later selects lacks necessary information should follow-up questions be asked as needed.
At this stage, only collect the business card; do not ask about the output path and do not start image generation.

Step 2: Then ask the user what images and sizes to generate

After the required items of the business card are complete, explain the six optional image categories to the user and ask about this session's mode:

Single: generate only 1 image of the specified type.
Multiple: generate only the type and quantity named by the user.
Full set: 1 image of each of the six categories by default, but the user is allowed to remove items, increase or decrease the quantity of the same category, or adjust the order.

Accept natural language, such as "make one white background image," "one selling point image and one detail image each," or "generate a full set." Do not force the user to reply by number.

When asking about the image type, you must also remind about the size:

When the user does not specify, white background images, display images, selling point images, and detail images are generated by default at 2000×2000 px、1:1; hero images and scene images are generated by default at 2000×1500 px、4:3。
The user may have all images in this round use the same size, or specify a size separately for each image.
If the user only states an aspect ratio but not pixels, first provide suggested pixels and ask the user to confirm; do not guess silently.
When the full set mode does not specify sizes, hero images and scene images default to 2000×1500 px、4:3, and the other four types default to 2000×2000 px、1:1。

Use the following wording, and do not wait until the generation planning stage to explain the default sizes for the first time:

Which type of image would you like to generate this time? You can choose white background image, display image, selling point image, hero image, detail image, scene image, or you can choose multiple images or a full set. Please also tell me the image sizes you need; if not specified, hero images and scene images will be generated by default at 2000×1500 px, 4:3, and the other types will be generated by default at 2000×2000 px, 1:1.

When the user has not previously provided any competitor links, ASINs, reference images, or personal visual ideas, you must, in the same round, ask whether they need you to search online for references:

You have not yet given me benchmark references. Would you like me to search online for excellent image sets in the same category as references? If needed, I will first break down excellent examples and then design; if not needed, I will design directly based on category experience and a unified visual concept.

For the boundaries of the six types, see references/image-types.md。

Step 3: Confirm the output location and naming

After the image type is determined, ask the user where they want the images saved after generation.

User-specified directory: use the user's directory, and first show the subdirectories and file naming scheme to be created.
If the user does not specify or replies "you decide": use references/output-naming.md the default directory in
Multiple images of the same product must be placed into subfolders by type (01-Product Main Image, 02-Display Image, 03-Selling Point Image, 04-Hero Image, 05-Detail Image, 06-Scene Image), and must not be mixed in the same directory; for the rules, see references/output-naming.md。
Images must not be generated before the output location is determined.
The directory name and the file name of each image must be written into the plan before image generation, so the user can understand what each image will be saved as.
Any rework must increase the version number; overwriting old images is prohibited.

Step 4: Establish the product fact lock and master control reference image

Lock the following facts from the business card and the user's images:

Product appearance, color, shape, structure, and material.
Labels, Logo, packaging text, and brand recognition points.
Actual selling quantity, packaging, and accessories included in the order.
Size, capacity, ingredients, functions, and selling point evidence explicitly provided by the user.
Brands, expressions, elements, and props that are prohibited from appearing.

Do not invent functions, sizes, ingredients, certifications, effects, or accessories based on common sense or reference images. Information that cannot be confirmed must be marked as "missing," and the image must be adjusted; do not guess.

Before generating display images, selling point images, hero images, detail images, or scene images, there must be a master control reference image that can stably lock the product appearance:

If the user already has a clear white background image: ask the user to confirm, then lock it.
If the user has clear multi-angle product images but no white background image: explain that a white background master control image needs to be established first; after obtaining the user's consent, generate and confirm it first, then continue.
If there are no product images at all: stop image generation and guide the user to upload them.

White background main image acceptance gate

When this round includes a white-background image or any secondary image, after the user confirms the entire set of plans, the white-background main image must first be generated separately, shown to the user for acceptance, and an explicit wait for a reply must be made. Before the user confirms the white-background image, it is forbidden to generate display images, selling-point images, hero images, detail images, or scene images.
After the user confirms the white-background image, lock that version as the product master control image. After that, the remaining planned types are generated, quality-checked, and saved continuously by the Agent, without pausing to ask the user for instructions image by image; after all are completed, the entire set of results is shown at once for the user to make a unified decision.
When the user generates only a single white-background image, the user must likewise be asked for acceptance. When this round does not include a white-background image but there is already one explicitly confirmed white-background master control image, the existing master control image may be used directly.

Reference images are only used to draw on composition, color tone, atmosphere, and layout; they must not change the product's true appearance, nor may they bring in competitor brands.

The fact lock only restricts verifiable product facts; it does not restrict reasonable commercial creativity. When optional fields in the business card are left blank, this must not cause degradation into mechanically applying a template, nor should the user be repeatedly questioned about every non-factual choice. Target audience assumptions, usage emotions, scenes, environmental props, colors, camera language, visual metaphors, layout, and marketing copy should be proactively completed by the Agent based on the category, confirmed selling points, target site, and this round's image task; these completions must not imply unverified product functions, certifications, materials, quantities, or order contents.

White-background image sole subject rule

By default, the white-background image shows only one complete, assembled sale subject. When detachable or visible parts such as the cup lid, tea filter, sealing ring, and inner liner are part of the product structure of that subject, they remain correctly assembled by default, and are not proactively disassembled, duplicated, or displayed side by side.
Only when the user explicitly requests displaying accessories, or when the product itself is a multi-piece set, combined pack, or set package that must be shown together to accurately express the sale quantity, does the white-background image show multiple independent items.
"The order includes a certain part" does not automatically equal "the white-background image must disassemble that part for display." When planning the white-background image, first determine whether it is an independently sold item, a set member, or a component part of the complete subject.
When it cannot be determined, prioritize using a sole complete subject and clearly state this in the plan; do not make the image look like an extra set of accessories is included just to show structure.

Step 5: Establish a unified visual concept and creative motif

Before generating any secondary image, first read and execute references/creative-direction.md. The entire set of images must not be merely a collage of six type templates; a unified but not rigid visual concept must first be formed.

At the same time, combine the visual anchors, reference image relationship assertions, dual references, layout baselines, and scale reference object rules of references/reference-image-techniques.md to control cross-image consistency and visual refinement.

The unified visual concept is generated by fusing independent creativity with insights from competitor/internet examples; the patterns from examples are only used to calibrate information architecture and layout rhythm, and the final image must be an expression exclusive to this product.

Two driving paths, branched according to whether the user provides references:

The user did not provide a reference (and step 2 also did not require finding references online): prompt = original meta-prompt text (references/meta-prompts/, verbatim consistent with the Feishu document) + unified visual concept + business card. This is the default path and does not depend on competitors.
The user provided a reference (ASIN, link, or reference image) or agreed to find references online: first follow references/competitor-benchmark.md Deconstruct, integrate the competitor style patterns into the original meta-prompt, then merge with the unified visual concept and business card. Competitors only contribute style patterns and do not replace the meta-prompt structure.
The Agent internally proposes at least 3 clearly different creative directions, compares product fit, suitability for the target site, selling point expressiveness, and image set scalability, then autonomously selects the strongest direction; by default, it does not expose all internal drafts to the user.
The selected unified visual concept includes at least: target audience hypothesis, buyer emotional value, core creative statement, brand temperament, color system, material and environment language, photography and lighting language, copy tone, and the visual motif running through the image set.
Before planning each image, the unified visual concept must be frozen as a "base tone contract," specifying at least: overall light-dark tendency, warm-cool color tone, primary and secondary colors, accent colors, color grading, font family, copy tone, and product texture. The base tone contract is only responsible for making the entire set look like it belongs to the same brand; it does not lock down each image's specific scene, background material, people, props, camera position, composition, or lighting direction.
When the user specifies composite styles such as "sporty white-collar," "natural technology," or "professional warmth," allocate the expressive focus each image bears across the full plan; do not require every image to cram in all style elements at once. Hero images and lifestyle images can carry people and emotion, display images establish product character, selling point images explain benefits, and detail images prioritize obeying the meta-prompt task of precision, cleanliness, and clear information.
White background images prioritize factual accuracy and platform compliance, and may not inherit the scene style of secondary images; display images, selling point images, hero images, detail images, and lifestyle images must share the same brand world, but each carries a different narrative and must not repeat the same layout.
When the user has not provided target users, scenarios, or visual preferences, the Agent should proactively make evidence-based business assumptions and mark them in the plan as "creative completion," and must not stall on a template combination of generic gradient background, centered product, and left-right text just because optional fields are empty.

Step 6: Output the generation plan and wait for confirmation

Before generating images, first show the selected unified visual concept, then provide the per-image planning table. Each image includes at least:

When presenting the unified visual concept, you must first explain to the user in one paragraph: what this round's concept design is, what the unified style is, and what role each image plays within this concept.

When the user has provided benchmark references or personal ideas, you must explain which parts of the concept come from user input and how they were integrated into the unified style.
When the user has provided no reference at all, you must be even more proactive in explaining where the concept comes from: what roles category common sense, online example insights, and autonomous creativity each played.
Do not just throw out the per-image planning table without explaining the concept design; when the user objects to the concept, revise the concept first, then the plan.

Item	Required content
Image type	One of the six categories
Commercial task	Which buyer problem this image solves
Copy	The actual text that will appear in the image; if there is no text, write "None"
Image content	Product, packaging, people, props, and environment
Composition and lighting	Subject position, lens, negative space, and lighting
Factual basis	From the business card or which reference image
Creative completion	Audience, scenario, props, visual metaphors, or benefit-point expressions proactively supplemented by the Agent
Output image size	Default 2000×2000 px, 1:1
Output file name	Full file name, not just the type

The user can modify any item. Only after the user explicitly replies with "confirm," "start generation," or an equivalent meaning may the image generation capability be invoked.

Step 7: Select the image generation backend

Select according to the following priority; do not reverse it:

Prioritize the built-in image generation capability of the current environment. If Codex, ChatGPT, or the current Agent already provides image_gen or an equivalent built-in image generation tool, call that capability directly. Do not ask the user to configure an API Key, and do not bypass the built-in capability in order to use a script.
Use the API fallback only when the built-in capability truly does not exist. Read references/image-backends.md, call scripts/generate_image.py。
A built-in tool temporarily failing, a network timeout, or a single generation failure does not mean "does not have image generation capability." First correct or retry according to the original tool's error message; do not silently switch to a potentially paid API.
Before using a paid API, show the provider, model, and estimated number of images to generate; if the user has not explicitly authorized use of that API in this round, obtain confirmation first.
API Keys are read only from environment variables; writing them into SKILL.md, scripts, prompts, logs, file names, or the final reply is prohibited.

Step 8: Generate one by one according to the selected types

First read the creative direction rules references/creative-direction.md and the unified prompt engineering rules references/prompt-engineering.md, then read only theoriginal meta-prompt textcorresponding to the types selected this time (full version, verbatim identical to the Feishu document "E-commerce Image Generation - Meta-Prompt Engineering"); do not load unselected types all at once, and do not read references/prompts/ the old simplified skeletons under it:

White background image:references/meta-prompts/01-white-background.md
Display image:references/meta-prompts/02-showcase.md
Selling point image:references/meta-prompts/03-selling-point.md
Hero image:references/meta-prompts/04-hero.md
Detail image:references/meta-prompts/05-detail.md
Scene image:references/meta-prompts/06-lifestyle.md

references/prompts/*.md is deprecated and is for historical reference only; it is prohibited as a source of image generation prompts.

Before generating secondary images, also read references/reference-image-techniques.md, and incorporate the visual anchor, scale reference, layout baseline, and lighting specifications into each prompt.

Execution requirements:

Use the confirmed product master control image as the product reference for each secondary image.
The first secondary image that passes user confirmation or Agent quality inspection can be locked as the "tone master control image." Starting from the next secondary image, when the image generation tool supports multiple reference images, the product master control image and the tone master control image may be attached at the same time: the former locks only the product appearance, and the latter locks only the basic light and shade, warm/cool tone, brand color scheme, color grading, fonts, and product texture; copying the specific scene, background material, people, props, or composition of the tone master control image is prohibited.
The meta-prompt is an "image generation prompt generator," not an image generation template. First read the references/meta-prompts/*.md full meta-prompt for the corresponding type above, put yourself into the role of the "visual strategist" in the meta-prompt, and according to its task constraints, output rules, formulas, and reference examples, combine the business card fields (product white background image, product name, brand name, product features, selling point hot words, target users, negative words, generation language, ratio), unified visual concept, and product fact lock,Produce a brand-new set of image generation prompts for this product. The output must conform to the formula and required fields of the meta-prompt for that type; the example product content is for format reference only and must not be mechanically reused. When the user provides an external prompt source (such as a multi-dimensional table record storing prompts), use the prompt in that source corresponding to this product's type verbatim, character by character, and only supplement the reference image consistency assertion, negative constraints, and output dimensions.
When generating images, use only the image generation prompt produced in step 3; do not feed the meta-prompt text itself directly to the image generation model, and do not use the meta-prompt as a skeleton to fill in and then generate images directly. All required fields of the meta-prompt must be reflected in the output, but the output is a complete executable image generation prompt for this product.
For different categories, proactively remove irrelevant modules and add modules that affect realism. For example, food focuses on ingredients and packaging quantity, electronics focus on ports and compatibility, and clothing focuses on fabric and wearing state.
The prompt for each secondary image must simultaneously carry the product fact lock, the base tone contract, and the unified visual concept for this round. The meta-prompt for the current type has priority over scene, layout, window, people, composition, and commercial task; the base tone contract serves only as a color and brand temperament overlay layer and must not mechanically copy the specific scenes of other images into it.
Translate the product features given by the user into buyer benefits and a visual story; do not merely lay out the business card fields onto the image as-is.
The entire set of modes must also be generated and quality-checked one by one, not six images generated in parallel batches.
After each image is generated, immediately quality-check and save it, then proceed to the next one.
By default, generate only 1 initial draft per image type. If the initial draft does not pass user acceptance or Agent quality check, modify only the failed items; a maximum of 2 revisions are allowed per type, for a total maximum of 3 generated images; record internally as "initial draft, revision 1/2, revision 2/2", with corresponding version upper limits typically v03. If revision 2/2 is reached and it still does not pass, image generation for that type must stop, and the failed item must be reported during delivery of the full set; do not continue consuming image generation attempts on your own.
Hero images, scene images, and other key images must also not be generated in multiple candidates by default. Additional generation is allowed only when the user explicitly requests multiple candidates, or when the plan approved by the user before image generation specifies the number of candidates; whether additional candidates count toward the revision count follows the plan description, and if unspecified, they count.
Except for the white-background main image acceptance gate, do not pause image by image during secondary image generation to wait for user decisions. The Agent performs its own quality check within the upper limit of at most 3 images per type, continues completing the other planned types, and finally presents the entire set of results to the user at once and asks the user to make a unified decision.
When the user only requests prompts, output only the corresponding prompts and do not call the image generation capability.

Step 9: Quality Check and Save

For each image, at least check:

Whether product shape, color, structure, label, quantity, and accessories are consistent.
Whether nonexistent functions, ingredients, certifications, dimensions, effects, or props appear.
Whether text spelling, language, numerical values, and readability are correct.
Whether the product is obscured by text, people, or props.
Whether it conforms to the commercial task of the selected image type.
Whether the generated image pixels, ratio, and safe margins conform to the plan.
Whether competitor brands, watermarks, or unrelated products appear.
Whether the secondary images belong to the same brand world while having different composition and narrative tasks.
Shrink the current secondary image and the master tone-control image and check them side by side: whether the overall light and dark, warm and cool tendency, brand color scheme, accent colors, font family, and product texture belong to the same brand. Different scenes, background materials, people, props, and compositions are allowed and encouraged, and failure must not be judged solely because the scenes differ.
Whether the current image has prioritized completing the commercial task specified by the meta-prompt for that type; if, for the sake of "unified style," it repeats other images' scenes, pedestals, runways, offices, or information modules, causing the image to be cluttered or the type boundaries to weaken, it must be redone.
Whether the composite style specified by the user is fully expressed at the level of the entire image set and carried by different images with divided responsibilities; it is not required that every image simultaneously show all style evidence.
How many revision opportunities the current type has already used; after reaching "revision 2/2," no further automatic rework or additional candidates may be made.
Whether there is mechanical template repetition such as "product centered + text on the left and right + ordinary gradient background," or simply turning business card fields directly into large text.
Whether the Agent's creative completion improves buyer understanding and emotional value while not crossing the boundary of product facts.

When it does not pass, rework only the failed items, with the version number incrementing and observing the total upper limit of at most 3 images per single type. Once the white-background image passes, immediately lock it as the master image; after the remaining secondary images are completed, save according to references/output-naming.md and, after the entire set is fully completed, report to the user all at once:

The absolute path of the output directory.
The absolute path and image type of each final image.
Which version was confirmed, and which fields are still items for the user to verify.

Default opening

When the user has not yet submitted a business card, directly use the concise opening below, without asking in advance about the image type or output path:

I will first learn about the product. Please send me the product business card; the product image, product name, and product category are required, and other information can be left blank if unknown. After the business card is confirmed, I will then ask you which type of image you want to generate this time, and where to save it after generation.