Ecommerce Image Asset Generator

It helps answer the higher-value questions first:
- **Which image assets matter most for this product or campaign?**
- **What should each image do in the funnel?**
- **What message should the image communicate?**
- **What should stay fixed vs change when editing an existing image?**
- **How do we turn business context into a usable visual brief or provider-ready prompt?**

## Solves

Ecommerce teams often fail visually in predictable ways:
- they make assets in the wrong order;
- they create hero images when trust images are the bottleneck;
- they generate attractive images that do not help CTR, CVR, or buyer understanding;
- design briefs are vague, so output quality becomes inconsistent;
- teams do not separate awareness images from conversion images;
- image edits drift too far from the original and break brand or listing consistency.

Goal:
**First identify the highest-value image assets for the business objective, then generate or edit only what is worth making.**

## Use when

Use when the user needs ecommerce visuals that serve a real commercial purpose, not just general inspiration.

Typical cases:
- launching a new product and deciding which visuals to create first;
- building a PDP, TikTok Shop, Shopify, Amazon, or marketplace image set;
- turning product notes into image briefs for design or AI generation;
- planning promo assets, offer visuals, or sale banners;
- generating new product or campaign images from prompts;
- editing existing images while preserving important elements;
- prioritizing a smaller, conversion-focused visual system instead of random one-off assets.

## Do not use when

Do not use this skill when:
- the user only wants a broad brand moodboard with no business goal;
- the task is video generation rather than image planning/editing;
- there is too little product context to define useful asset priorities;
- the user wants a full brand identity system instead of ecommerce assets;
- provider execution is requested but auth / endpoint / runtime details are unavailable.

## Core principle

**Plan before generate.**

The strongest ecommerce image work does not start with the model.
It starts with the conversion job.

Every asset should map to one of these jobs:
- stop the scroll
- explain the product faster
- reduce doubt
- make the value clearer
- increase trust
- improve click-through or conversion
- support an offer, launch, or promo moment

## Working modes

### Mode 1: Plan-only
Use when no provider is configured, or when the user wants to approve direction first.

Return:
- asset priority list
- brief for each asset
- overlay / copy suggestions
- prompt drafts for later generation

### Mode 2: Plan + Generate
Use when a supported provider is available and the user wants images created.

Return:
- asset plan
- provider-ready prompts
- generated outputs or saved paths / URLs
- notes on what to improve next

### Mode 3: Edit Existing Image
Use when the user already has source images and wants controlled changes.

Return:
- what must stay fixed
- what should change
- edit brief / prompt / payload plan
- edited outputs or saved paths / URLs

### Mode 4: Batch Asset Set
Use when the user needs a coordinated visual set for a PDP, campaign, listing, or launch.

Return:
- prioritized asset set
- image-by-image brief
- generation order
- prompt set / payload guidance

## Inputs

Ask for the minimum useful business context:
- product name and category
- target audience
- platform / channel
- business goal (CTR, CVR, education, trust, promo, launch, retargeting)
- key selling points
- strongest proof points
- offer / promo context, if relevant
- existing assets, if any
- compliance boundaries or prohibited claims
- provider choice, if generation/editing is requested
- source image(s), if editing is requested

## Workflow

### Phase 1: Clarify the visual job
1. Restate the product, audience, and commercial objective.
2. Identify what buyers still do not understand or trust.
3. Decide which image types would solve that fastest.
4. Prioritize the assets instead of listing everything.

### Phase 2: Recommend the right asset mix
Choose only the most commercially useful image types, such as:
- hero image
- feature / benefit image
- comparison image
- problem-solution explainer
- trust / review image
- offer / promo image
- lifestyle / use-case image
- marketplace infographic image
- UGC-style image concept

### Phase 3: Brief each asset clearly
For each asset, define:
- asset type
- business objective
- key message
- visual direction
- on-image text / overlay guidance
- must-show product elements
- compliance / avoid notes

### Phase 4: Generate or edit if supported
If a provider is configured:
1. Convert the brief into provider-ready prompts or payload logic.
2. Generate or edit the asset(s).
3. Review whether the output actually matches the commercial job.
4. Suggest the next iteration.

If provider details are missing:
- stop at the planning + prompt layer;
- do not fake execution.

## Provider model

Treat generation as **provider-pluggable**.

Preferred flow:
1. Plan assets first.
2. If execution is requested, ask which provider should be used.
3. Ask only for the minimum provider-specific info.
4. Route generation/editing through that provider.

Recommended provider order:
- **Seedream 5.0 via ARK API** for a modern API-first image flow
- **Nano Banana Pro / Nano Banana 2** for strong text-to-image and image edit workflows
- **Jimeng 4.0** for Chinese-first prompting, grouped outputs, and multi-image workflows
- **Other compatible image APIs** when the user provides endpoint + auth + required parameter mapping

If provider details are missing, remain in **Plan-only** mode.

## What to ask when execution is requested

Ask the smallest useful set, not a giant questionnaire.

### Universal questions
- Which provider should be used?
- Is this text-to-image, image edit, or batch asset generation?
- How is access configured? (env var, API key, token, script, existing runtime)

### Provider-specific minimums
#### Seedream 5.0 / ARK
Ask for:
- whether `ARK_API_KEY` is configured;
- preferred model, if not default;
- desired output size, if important.

Default assumptions when not specified:
- model: `doubao-seedream-5-0-260128`
- size: `2K`
- response format: `url`
- watermark: `false`

#### Nano Banana
Ask for:
- which version / workflow to use;
- whether key/runtime is configured;
- whether there is an input image for edits.

#### Jimeng 4.0
Ask for:
- auth/signing method;
- whether this is text-to-image, edit, or grouped generation;
- whether grouped outputs or `force_single` are needed;
- whether input image URLs are already available.

## Output

Return a practical asset package:

1. **Asset priority list**
2. **Per-asset commercial brief**
3. **Visual direction + overlay guidance**
4. **Provider-ready prompt or edit plan**
5. **If execution runs:** file paths, URLs, or iteration notes

## Quality bar

A strong output should:
- plan before generating;
- map each image to a business objective;
- distinguish awareness images from conversion images;
- prioritize a smaller, better asset set over a generic long list;
- avoid misleading or non-compliant claims;
- preserve critical elements during edits;
- stop honestly at the planning layer if provider execution is unavailable.

## What “better” looks like

Good output should make it obvious:
- which image to make first;
- why that image matters;
- what that image should say visually;
- how to brief or generate it without ambiguity;
- what to improve next if the first version is weak.

## Resource

See `references/output-template.md` and the provider notes in `references/`.