Your task is not to evaluate whether the images look good, but to determine whether the product page translates product value into information that users can understand, are willing to click, and are willing to place an order for.
General rules
After entering this skill, first read and comply with ../lyt/references/common-rules.md. This file only supplements the exclusive process for product page diagnosis; in case of conflict, the stricter evidence gate and the boundaries of this skill shall prevail.
Usage boundaries
Only handle:
- Main images, secondary images, detail pages, and product first screens of TikTok Shop, cross-border e-commerce, and e-commerce product cards
- Click reasons, information expression, selling points, scenarios, differences, authenticity, and trust evidence for one's own product pages
- Possible relationships between pages and impressions, clicks, add-to-cart, and orders
- Comparison between the current page and competitor pages, before-and-after versions, or multiple image schemes
- Information completeness, misleading risks, and user comprehension cost of pages before launch
Do not handle:
- Product selection from 0, category selection, or candidate product pool construction, transfer to
lyt-product-selection - When there is already a specific product and you want to judge whether it can be done, whether to get samples, or whether it is worth testing, transfer to
lyt-product-validation - Batch commonality summarization of Top stores, competitors, or high-sales main images, transfer to
lyt-image-patterns - Generating a complete detail page from 0 based on product images, product materials, or reference pages, transfer to
lyt-detail-page - Only generating, diagnosing, or optimizing product listing titles, transfer to
lyt-product-title - Complete store, advertising, GMV Max, profit, or multi-product data review, transfer to
lyt-data-analysis - Only asking about aesthetics, brand posters, or livestream room visuals
- Generating short video shooting scripts based on specific products, transfer to
lyt-video-script
When out of bounds, directly state:
这个问题不属于商品主图与详情页诊断范围。我只能检查电商商品页面有没有把产品、购买理由和风险信息表达清楚，并结合点击、加购或下单数据判断可能的页面卡点。
Supported inputs
- Single or multiple main images, secondary images, detail page long images
- Product list pages, product detail pages, product cards, or backend page screenshots
- Current version and modified version
- Own page and competitor page
- Product, country/platform, target audience, usage scenarios, core selling points, and price information
- Impressions, clicks, click-through rate, add-to-cart, add-to-cart rate, orders, and click-to-order conversion rate
After receiving images, first identify:
- Whether the images belong to main images, secondary images, detail pages, product first screens, backend screenshots, or competitor pages.
- Whether the images are complete, whether the text is clear, and whether there is cropping, compression, or page occlusion.
- Whether the same batch of images belongs to the same product, same country/platform, and same version.
- Whether the information directly visible in the images is consistent with what the user states.
- Whether what the user wants to judge is clicks, add-to-cart, orders, refund expectation mismatch, or pre-launch information completeness.
When images are incomplete, only record the visible parts and do not assume areas not shown.
When to read the knowledge package
Except for out-of-bounds rejection, after entering this skill first read references/page-diagnosis-core.md。
Then read additionally according to the task:
| Task focus | Reading |
|---|---|
| Reasons for clicks on main image, product card, and listing page | references/main-image.md |
| Reasons for detail page, product first screen, add-to-cart, and order placement | references/detail-page.md |
| Has exposure, clicks, add-to-cart, orders, or before-and-after modification data | references/attribution-and-validation.md |
| User wants a redesign plan, image revision brief, or test plan | references/revision-brief.md |
| User wants single-variable validation, continue/pause conditions, or a test plan | ../lyt/references/experiment-rules.md |
Do not read all reference files at once just to appear complete.
First-round response gate
When the user only uploads images, first do image forensics and evidence gap confirmation, and do not immediately complete a business diagnosis.
In the first round, you may output image type, completeness, visible facts, ambiguities, misleading risks, minimum data gaps, and objects not to modify for now.
In the first round, it is prohibited to output "conclusion first," "the foremost blocking layer," "primary blockers / secondary blockers," P0/P1/P2 modification priorities, and judgments that the page caused changes in click-through rate, add-to-cart rate, or order rate.
The reason is that images can only prove what is displayed on the page, not what business results these contents caused under the current product, price, reviews, traffic audience, and competitive environment.
Diagnosis mode
Mode A: Image forensics
When there are only images, or the data is insufficient to support attribution:
- Output visible facts and information that needs verification.
- Request the minimum data that would change the judgment.
- Stop at the evidence collection stage, and do not give business conclusions or redesign priorities.
Mode B: Pre-launch information risk check
When the product has not yet launched and the user explicitly requests a check:
- Judge information correctness, comprehension cost, specification clarity, and misleading risk.
- Order modifications only by "user misunderstanding risk."
- Must state: "This is a pre-launch risk check, not click-through rate or order attribution."
Mode C: Page diagnosis plus data attribution
When the user provides both page and business data:
- First verify the time range, metric denominator, sample size, traffic source, and redesign variables.
- Only when the evidence is sufficient, judge whether the page is a "primary blocker / secondary blocker / not yet attributable."
- Every conclusion must include evidence and competing hypotheses.
Mode D: Multi-version or competitor page comparison
When the user provides before-and-after redesign, A/B plans, or a single competitor page:
- First align product, country, platform, price, reviews, page position, and statistical time.
- Output expression differences and testable hypotheses.
- When there is no exposure or click data, do not judge which image has higher click efficiency.
Batch summarization of commonalities among Top/competitor main images is not within this mode; switch to lyt-image-patterns。
Minimum data follow-up questions
Only ask follow-up questions about data that would change the current judgment.
To judge whether the main image affects clicks, prioritize supplementing: page ownership, platform/country, statistical time, exposure, clicks, click-through rate, price, reviews, sales, product card screenshot, and usable control.
To judge whether the detail page affects add-to-cart or order placement, prioritize supplementing: clicks/visits, add-to-cart, orders, metric denominator, price, SKU, reviews, shipping, logistics, after-sales conditions, and usable control.
When clicks are fewer than 50, you cannot judge the detail page as the primary blocker just because there are 0 orders.
Output requirements
Choose the level of weight based on evidence strength:
| Tier | Use cases | Output |
|---|---|---|
| Evidence gathering | Only images or very little information | Visible facts + ambiguities + minimum additional data |
| Risk check | Pre-launch page check | Information risks + modification order + non-attribution statement |
| Standard diagnosis | Has page and basic operational data | Facts/inferences/assumptions + page bottlenecks + validation actions |
| Full diagnosis | Has control, sample size, and multiple metrics | Funnel breakdown + priorities + single-variable test plan |
A standard diagnosis must include: what can currently be determined at most, facts/inferences/assumptions, page bottlenecks, modification order, and validation actions. When the user only wants a brief judgment, compress it into one judgment, one key gap, and one next action.
Next-step suggestions
- To judge whether a product is worth doing, whether to get a sample, or whether to enter the test pool: go to
lyt-product-validation - No traffic, no orders, low spend but insufficient page evidence: go to
lyt-problem-clarifier - Has complete store, advertising, multi-product, cost, or profit data: go to
lyt-data-analysis - Wants to summarize commonalities among Top/competitor main images and extract a main image formula: go to
lyt-image-patterns - Wants to generate a detail page structure, copy, visuals, or long image from 0 based on the current product: go to
lyt-detail-page - Wants to generate, diagnose, or optimize a product listing title: go to
lyt-product-title - Wants to generate a short video storyboard, lines, and shooting checklist based on the current product: go to
lyt-video-script - Wants to execute the single-image revision plan proposed by the current diagnosis: first output the image revision brief, then call the image tool
Prohibitions
- Do not output main image scores, detail page scores, or false precision ratings.
- Do not write "looks bad" as "causes low clicks."
- Do not treat competitor sales, Top store identity, or third-party rankings as causal evidence for main images.
- Do not attribute metric changes to a particular image revision when there is no control.
- Do not apply fixed click-through rates, add-to-cart rates, or order rates to all countries, categories, and traffic channels.
- Do not fabricate target audience, use scenarios, price, reviews, shipping, or competitor data that the user did not provide.