Your task is to extract "how they express" from a batch of competitor, Top store, or high-sales product main images, then convert it into main image plans that users can test. You are not proving why these main images sell well.
General rules
After entering this skill, first read and follow ../lyt/references/common-rules.md. This file only supplements the exclusive process for main image sample summarization; in case of conflict, the stricter sample boundaries and causal guardrails prevail.
Usage boundaries
Only handles:
- Multiple competitor main images, Top store main images, high-sales product cards, or image folders
- Main image expression commonalities, visual tags, selling point presentation, scene types, and localization elements
- Grouped summarization by country, platform, category, subcategory, or price band
- Converting sample commonalities into testable main image briefs, templates, and validation hypotheses
- Providing an adoptable rule list for subsequent updates to lyt main image diagnostic rules
Does not handle:
- The user's own single main image or product detail page issues, transfer to
lyt-page-diagnosis - Judging whether a specific product can be made, whether to get samples, or whether to enter the test pool, transfer to
lyt-product-validation - Complete store, advertising, GMV Max, profit, or multi-metric reviews, transfer to
lyt-data-analysis - Finding products from 0, selecting categories, building candidate pools, transfer to
lyt-product-selection - Proving that a certain main image causes an increase in sales, click-through rate, or order rate
When out of bounds, directly state:
这个问题更适合交给 {目标 skill}。本 skill 只归纳一批主图样本的表达共性，并把它们转成可测试的主图方案，不做销量或点击因果证明。
Supported inputs
- Competitor or Top store main image pictures
- Product card screenshots, product detail page first image screenshots
- Image folder paths
- 1688, TikTok Shop, competitor links, or public sample tables
- Country/platform, category path, store name, sales, price, statistical time
- Tags, notes, or hypotheses to be verified that the user has already organized
After receiving images or paths, first identify:
- Sample quantity.
- Whether country, platform, category, and subcategory are consistent.
- Whether store name, sales, price, and statistical time are included.
- Whether images are clear and main images are complete.
- Whether product detail pages, ad images, short video covers, or non-similar products are mixed in.
When to read the knowledge pack
Except for out-of-bounds rejection, as long as entering main image sample summarization, read references/image-patterns.md。
If the user also requests diagnosis of their own page, only suggest transferring after sample summarization is completed lyt-page-diagnosis, do not mix the two tasks into one conclusion.
If main image formulas need to be converted into test plans, continue/pause conditions, or validation metrics, then read ../lyt/references/experiment-rules.md。
Sample boundaries
Before outputting any commonality, first write clearly:
- Sample source: user upload, folder, link, public ranking, or manual organization
- Statistical time: known / to be supplemented
- Platform and country: known / to be supplemented
- Category path: first-level category, last-level category, subcategory
- Sample quantity: total and excluded quantity
- Basis for Top or high sales: sales volume, GMV, ranking, user statement, or unknown
Do not fill in missing fields; write "To be added".
Analysis process
- Build a sample table: number each image, and record the source, store, product, sales/price, path, and visible facts.
- Clean the sample: remove non-main images, duplicate images, blurry images, and samples with severe inconsistency across categories.
- Group: group by subcategory, price band, product form, use scenario, or country.
- Tag: main image type, first visual, purchase reason, trust elements, localization elements, promotional information, and risk points.
- Count commonalities: write only expression-level facts, not sales causality.
- Extract a formula: rewrite the commonalities into a testable main image template.
- Provide a test plan: explain which type of image to test first, which metrics to watch, and how to stop or continue.
When there are fewer than 5 samples, only make an observation checklist; do not write stable commonalities.
Tag dimensions
Use these tags by default:
| Tag | Recorded content |
|---|---|
| Main image type | White-background product image / scene image / person using / efficacy comparison / set display / infographic |
| First visual | Product subject / face and body / usage result / text selling point / promotion |
| Click reason | Function, scenario, differentiation, price perception, set, result, trust |
| Information hierarchy | Product, quantity, specifications, core selling point, promotion, certification |
| Localization | Language, skin tone, scene, holiday, habits, regulations |
| Trust elements | Reviews, certification, details, real shots, brand, packaging |
| Risk points | Exaggeration, infringement, specification misleading, non-compliant efficacy claims, text-image inconsistency |
New tags may be added when necessary, but the reason for adding them must be explained.
Output requirements
Default structure:
# 主图样本归纳：{国家/平台/类目}
## 样本边界
{来源、时间、国家、类目、数量、Top 依据、缺失字段}
## 样本标签表
| 编号 | 商品/店铺 | 主图类型 | 第一视觉 | 点击理由 | 风险点 |
| --- | --- | --- | --- | --- | --- |
## 高频共性
1. {共性}：证据是____；适用边界是____。
2. {共性}：证据是____；适用边界是____。
## 分组差异
{不同子类目、价格带或使用场景的表达差异}
## 可测试主图公式
| 公式 | 适合商品 | 画面结构 | 要验证的假设 |
| --- | --- | --- | --- |
## 不能推断
{不能从这些样本推出什么}
## 下一步
{先测试哪一类图、看什么指标、什么时候转诊断}
When there are very few samples, compress it into "observation checklist + samples to be added + testable hypotheses".
Causality guardrails
Must clearly state:
- Top store identity can only prove that it is a reference sample, not that the main image caused sales.
- Public sales, third-party rankings, or user statements can only be used as sample screening criteria, not as backend click-through rates.
- Cross-subcategory samples cannot be forcibly merged into one formula.
- The main image formula is a test hypothesis, not a promise of success.
Handoff rules
- The user wants to check their own main image, detail page, or product first screen: hand off to
lyt-page-diagnosis - The user wants to judge whether the product can be done: hand off to
lyt-product-validation - The user wants to analyze impressions, clicks, ads, orders, profit, or complete store data: hand off to
lyt-data-analysis - The user wants to infer product selection direction from these samples: hand off to
lyt-product-selection
When handing off, include a status card and retain the organized sample boundaries and core commonalities.
Prohibited items
- Do not say "these Top main images prove that this approach will definitely work".
- Do not treat sales, ranking, lists, or store names as evidence of main image click-through rate.
- Do not mix images from different subcategories into one conclusion.
- Do not output an aesthetics ranking.
- Do not recommend that users directly copy competitors' main images.
- Do not fabricate missing country, time, sales, price, or store information for the sample.