Turn product materials into listing visuals for Taobao, JD

E-commerce Visual Planning and Execution

Full English companion to the canonical runtime entrypoint, SKILL.md. English and bilingual output must preserve the same evidence, compliance, routing, and execution rules.

Turn raw product materials into a visual plan a designer or image model can execute. Default to Chinese; switch language when requested without changing factual scope.

Core principles

1. Deliver only after the self-review passes: before the final draft is delivered, complete the five-part self-review in Step 6 and every dimension must score ≥80; if any dimension is below 80, rewrite first and never present an under-threshold draft as final. Self-review happens inside the same reply, is not a pause, and cannot be skipped for "one-shot" or "no interruptions" requests; attach a self-review record to the delivery.
2. Check evidence and legality separately: start with user-supplied product, packaging, specification and qualification materials for facts. Applicable mandatory law cannot be overridden by those materials or a user request. Evidence does not by itself make an advertising claim lawful; assumptions cannot fill factual gaps.
3. Route by intent: deliver the requested layer, without forcing a full workflow or asking for information already supplied.
4. Never turn assumptions into facts: do not invent qualifications, tests, patents, sales, reviews, rankings, efficacy, certifications, or platform rules.
5. Lock visual identity before execution: product identity, composition, palette, lighting, typography, background, reference fidelity, and negative constraints must stay consistent.
6. Treat platform rules as time-sensitive: verify official sources when current dimensions, image restrictions, category qualifications or claim policies matter. Label unverified requirements.
7. Do not provide final legal conclusions for food, supplements, medical devices, health care, children's goods or other high-risk categories. Check the target market and category.

The skill does not cover pure short-video/live-selling scripts, advertising or store-operations optimization, or final lawyer-level legal conclusions (risk screening is supported), and does not automatically publish listings.

Execution modes

The kickoff questionnaire (Step 0) is not a checkpoint: preference items are optional, sent once, and may use defaults; a missing name must still be resolved. Silence or skipping preferences is not authorization to invent a name. Self-review is not a pause either and is required in every mode.

Shared output contract: read references/output-contracts.md before a complete plan or formal audit in any mode, including a full Standard strategy/storyboard. Small isolated wording edits do not require the entire template.

Standard

Use when critical evidence or direction is missing, or the user asks for strategy/storyboard before execution.

1. Deliver the visual strategy, evidence ledger and evidence gaps together.
2. Deliver the supported storyboard; mark unresolved parts rather than inventing scene details.
3. Once the preliminary product-name check is resolved, pause once, after the Storyboard, combining necessary evidence requests and direction review. After confirmation, proceed to execution and quality gates without per-image approvals.

One-shot

Once the product name is supplied or image-based naming is authorized, if the user asks for a complete answer in one pass ("just do it", "no questions"), skip the kickoff questionnaire and complete strategy, storyboard, execution cards and quality gates, labeling presentation assumptions without inventing product facts. If One-shot was chosen only because inputs suffice, still send the kickoff questionnaire once (missing dimensions only), then complete everything in one pass once the user answers or skips. Self-review is still mandatory.

Audit / revision

For existing images, copy, prompts, or PDPs, identify problems in conversion logic, hierarchy, evidence/compliance, platform fit, image/copy alignment, and style drift. Make the minimum necessary changes.

Designer handoff

Use the shared output contract for cards that a designer or image model can execute without guessing.

Default mode (when the user does not specify)

Explicit mode, sequence and checkpoint requests take priority. Otherwise, first match wins:

1. Existing finished images, PDPs, copy or prompts to review/revise → Audit / revision. Source product/packaging photos alone do not trigger Audit.
2. A brief for a designer or operator to execute directly → Designer handoff.
3. Product, platform/market, audience, necessary assets and evidence suffice for this scope → One-shot. Do not require three selling points just to fill a quota.
4. Missing critical evidence, unclear platform/audience or direction → Standard, with one combined checkpoint after the Storyboard.

State the selected mode in one sentence. For explicitly requested One-shot or handoff with missing materials, complete supported parts and list blockers outside publishable copy without fabricating facts or implying unsupported work is complete.

The kickoff questionnaire applies to Standard, One-shot and Designer handoff; Audit asks only revision-relevant dimensions (platform, style, must-keep items). Skip it when the user says not to ask, or when the relevant dimensions are already clear from the message and materials.

Step 0 — Task routing and minimum inputs

Identify the requested surface: main images, PDP/detail page, Amazon Listing/A+/Brand Story, ad images/video covers, on-image copy, storyboard, image prompts, compliance review/revision, or cross-border localization.

Prioritize product identity, platform/market, audience, selling points/specifications, evidence, and source/reference images.

Confirm the product name at the start

• Record and use a supplied product name or an explicit usable category name. Do not ask again merely because a brand/series name is absent, or replace a known original without authorization.
• If no name is supplied and image-based naming is not authorized, ask before strategy/Storyboard: “What is this product called? You can provide a name, or let me propose one from the image.” With a clear image, offer one or two candidates labeled “image-based proposed name, awaiting confirmation”; if the category is unclear, ask rather than guess.
• Wait if the user will supply the name. If they choose Agent naming, use visible style/form plus the generic category, record “image-based proposed name, not an official original”, and continue without forcing another approval. If that authorization was already explicit, proceed without asking again. Do not add unverified materials, manufacturing, functions, origin or brands, or label a candidate as a confirmed original.
• Resolve the name or obtain naming authorization before a formal strategy/execution card; visible facts and evidence gaps can be organized while waiting. This preliminary question applies in all modes and is separate from Standard's post-Storyboard direction/evidence checkpoint. Do not defer a missing-name question until after the storyboard.

Kickoff questionnaire (optional; work proceeds without answers)

Once the name is supplied or image-based naming is authorized, send one preference questionnaire before starting. A missing-name question may be bundled into that round but requires an answer or naming authorization. Fuller answers give a better-fitted plan; if the user does not answer, replies "skip"/"just do it", or simply supplies materials, continue with the stated default assumptions, ask nothing further, and do not relax evidence or compliance boundaries.

• Ask only for dimensions missing from the message and materials; reuse existing project context.
• Send preference items together, all optional, each with its default; the user may reply with numbers only (for example "1 Taobao 3 women 25–35, rest default").
• If the environment has a choice-picker tool, use it for choice questions such as platform, style and scope; otherwise use numbered text. The template is in references/output-contracts.md.
• One round only. Once naming is resolved, after any complete or partial reply proceed without re-asking the same preference dimension; gaps that block compliance or product identity (missing certificates, missing product photos) are listed outside the copy, not asked again.
• Label default assumptions explicitly in the plan.

#	Dimension	Ask	Default if unanswered
1	Platform / market	Taobao/Tmall, JD, Pinduoduo, Douyin Shop, Xiaohongshu, WeChat Channels, Amazon, Shopify…	Infer from materials; else a Taobao/Tmall main-image set
2	Scope	How many main images, PDP, Listing/A+, image prompts needed?	About 5 main images, each with execution card and prompt
3	Audience & scene	Who buys, where, when	Infer from materials, labeled as assumption
4	Core selling points	The 1–3 points to lead with; which have certificates, tests or parameters	Distill top 3 from materials
5	Style & tone	Minimal premium, lifestyle seeding, direct promotion, national-style, tech…; references or competitors	Category and platform convention
6	Price & promotion	Positioning; any campaign, gift or bundle	No price or promotion messaging
7	Must keep / avoid	Brand colors, logo, fixed structure, banned words, unwanted elements	Product photos prevail; default Negative Constraints
8	On-image text	Post-production type, no added text at all, or a short model-rendered headline	Layout-reserved; no model-rendered text
9	Reference material	Product/packaging photos, opportunity or competitor analysis, brand guide	Text description only; no invented appearance

A selling point supplied verbally is logged as "user-stated". Without original documents or test evidence it remains Needs proof; being asked or answered does not upgrade it to Verified.

Distill reference material into hard constraints

When the user supplies opportunity analysis, competitor references, brand material or others' creative, distill supported hard constraints before writing (usually three to five; do not pad sparse materials) (audience, differences to stress, practices to avoid, structures worth borrowing) and list them in the plan. Learn structure (sentence patterns, information hierarchy, module roles), not surface (language, decorative elements, palette); never transplant others' facts, data or promises. Reading the material without visible use is a failure.

The no-copying restriction concerns competitors' or others' material; user-owned brand colors and authorized visual references explicitly marked to preserve still follow Reference Fidelity.

Apart from the preliminary name check, clarification of conflicting product materials and Standard's post-Storyboard evidence/direction checkpoint, do not ask further preference questions. Continue with labeled presentation assumptions for remaining display-level gaps; when a gap would change product identity or compliance boundaries, list it outside the copy and never assume product facts.

Step 1 — Build an evidence ledger

Every factual, numerical, efficacy or qualification claim intended for the final page enters the ledger first.

Claim	Source	Status	Final treatment
Directly supported by materials whose product/scope has been checked	Packaging image, specification sheet, specific document	Verified	Preserve meaning; a verbal assertion does not establish testing/certification
Benefit derived from a verified parameter	Same source	Supported	Translate value without expanding scope
Claim requiring a test, certification, rule or approval	Missing evidence	Needs proof	Use neutral wording, or delete the item and list it in a pending-confirmation list outside the copy
Fabricated, misleading or unsupported regulated claim	No lawful support	Disallowed	Exclude from final copy

Only Verified and Supported claims may appear as facts. Record a file/page, image ID or official URL, applicable SKU, market, conditions and date when relevant. An example image establishes its appearance, not the truth of its origin, certification, specifications or commercial claims. Gaps go in a pending-confirmation list outside the copy, never in image copy or the copy fields of execution cards.

If a claim is verbal and the original material has not been inspected, record that source honestly rather than saying the material was verified. A structure or material alone does not establish performance: "has feet/pads" cannot become "stable placement" or "non-slip" without supporting evidence. Calling this a "neutral scene description" does not remove the evidence requirement.

Inference is not fact. Sources fall into four types: documentary proof (test reports, certificates, official documents), supplier claims, visual observation and guesswork. Only documentary proof, and supplier claims already cross-checked against parameters or the real item, may enter finished copy; visual observation and guesswork stay in notes outside the copy. When a merchant claim conflicts with the real item, the item prevails and the user is asked. Write material, plating, synthetic and treatment attributes with the full standard name and its qualifiers, not shorter trade slang. Specification areas and parameter tables contain only Verified data.

These restrictions concern factual claims and do not cancel Step 0's authorized image-based naming. A working name combining visible form and category may appear in KV1, with its internal source still labeled “image-based proposed name, not an official original”. It proves no material, manufacturing, performance or qualification claim and does not upgrade observation to documentary proof.

Step 2 — Select the conversion driver

Choose one primary and at most one secondary driver.

Driver	Typical use	Visual emphasis
Visual / sensory	Beauty, jewelry, home, gift boxes	Materials, light, proportion, packaging
Pain point / efficiency	Cleaning, storage, tools	Use context, process, fewer steps
Rational / parameter-led	Electronics, appliances, equipment	Specifications, mechanisms, compatibility
Trust / evidence-led	Food, tea, baby products, high-value goods	Origin, craft, qualifications, traceable evidence
Emotional / identity-led	Fragrance, apparel, gifting, pets	Aesthetic, rituals, use scenes, identity
Price / bundle / promotion	Pinduoduo, Temu, Douyin, kits/refills	Included items, unit/bundle price, conditions, dates

Translate up to three supported purchase reasons for the requested scope with Feature -> Advantage -> Benefit -> Evidence; never invent extra reasons to fill three slots. Conversion pressure never justifies unsupported claims.

For price-led work verify currency, SKU, quantity, unit/bundle total, eligibility, thresholds, start/end dates, and shipping. If evidence is missing, show known bundle contents only; do not invent reference prices, discounts, lowest-price claims, inventory, timers, or gifts.

Step 3 — Lock the Campaign Style

Define:

• Product identity: immutable packaging text, logo, color, proportions, structure, materials and details.
• Composition: product coverage, camera angle, whitespace, reading direction, people/props.
• Palette: primary, supporting and accent colors.
• Lighting: natural/studio, hard/soft light and rim lighting.
• Typography/density: headline, numbers, body and mobile readability.
• Background/scenarios: plain, studio, real-world, or brand atmosphere.
• Reference Fidelity: what must stay and what may change.
• Negative Constraints: forbidden new logos, accessories, textures, anatomy, structure or efficacy symbols.

Use the real uploaded product/packaging as the source of truth. Decorative improvement must not change packaging, flavor, size, logo, accessories or structure.

Step 4 — Plan the storyboard

Do not hard-code five images. Five is a starting point; adapt to platform, user scope and actual information.

Role	Decision task
Hero / KV1	First identify what the product is; purchase reasoning is secondary and optional
Benefit	Translate a key feature into customer value
Proof	Establish trust with real evidence
Compare	Reduce selection friction using comparable facts
Scene	Show relevant use contexts
Spec / CTA	Explain variants, bundles, conditions or the next decision

KV1 product recognition: product before feature

The first product image answers “what is this?” in the order product identity → distinguishing form/style → at most one supporting purchase reason. Establish recognition before evaluating appeal; do not omit the product name to make room for a benefit.

• Use the name confirmed in Step 0 or an image-based proposed name authorized by the user; record its naming source, generic category and SKU. Use a clear known original within the copy limit without replacing it arbitrarily. If too long, retain an internal original-to-short-name mapping and preserve category plus essential identifiers without adding unverified materials or effects.
• Where text overlays are allowed, the headline itself names the product/category. A card label, evidence note or tiny caption does not satisfy this. Brand names, metaphors, scenes and features cannot substitute for nouns such as “necklace” or “storage rack”. Put feature/emotional wording in support; a product name alone is sufficient.
• Both original and proposed names identify the product first. Metaphors or structural descriptions alone cannot serve as the first-image headline; supported wording may become support or a later feature-image theme. Examples are illustrative, and the rule applies to all products; see examples/README.md. Translate the established identity faithfully in another output language.
• Show the product's overall relationship, identifying form and real proportions clearly; models, scenery and decoration support recognition. Distinguish color choices from bundle quantities; gold/silver appearance proves neither material nor a two-piece offer.
• If the user requests no added text or the placement disallows overlays, add no headline. Establish identity through the real product and original packaging, record its name in the card, and explain the visual recognition basis. Preserve packaging text without inventing labels; inspect the eventual image separately.
• KV1 recognition check: at mobile size, hide support copy and inspect only the product and any permitted headline. Can an unfamiliar viewer name the product category? If guessing is required, Fail and revise the headline/product presentation. This applies to the first product image; later Benefit/detail images may lead with features without repeating the name.

A PDP may follow: Opening -> Top Benefits -> Mechanism/material/craft -> Scenes -> Proof -> Selection -> FAQ -> Specs/CTA. Omit unnecessary modules.

Order information by the buyer's decision path, not by product construction: what is it → what is it to me → why this one → why trust it → how to buy. Each PDP module resolves one decision question, in the order doubts arise.

Load the relevant section of references/platform-playbooks.md for platform work. Keep Douyin Shop separate from TikTok Shop, and use the separate Xiaohongshu and WeChat Channels/WeChat Shop sections. Verify current rules using references/source-notes.md when dimensions, text permissions, claims or category requirements matter.

Step 5 — Execute

For each image/module deliver:

1. Visual task: the decision question it solves.
2. Scene/composition: subject, camera, background, materials, light and hierarchy.
3. On-image copy: one headline and necessary short support, checked against the copy specifications below.
4. Design rationale: first focus, reading order, emphasis and annotation/disclaimer placement.
5. Image-generation prompt: subject + composition + scene + materials + lighting + Style Lock + reference constraints + text rendering method.
6. Negative Prompt / forbidden changes: packaging distortion, incorrect logos/specs, new structure, text artifacts, anatomy errors and style drift.
7. Evidence notes for claims, numbers, qualifications or comparisons.

Image prompts go to an AI image model by default; do not ask who receives them. Every prompt must be self-contained: do not depend on earlier cards; state supported product structure/materials, lighting, composition, whitespace and forbidden changes. If source product images are needed, identify them and preservation requirements so the prompt and specified images can be handed off together. Do not invent structure or materials to fill fields; repeat the Style Lock in each prompt. State what the model must not generate: on-image text (by default), logos and original packaging text, certification or qualification marks, and specific numbers.

Ten-line minimal execution card. Teaching inputs: a structure drawing and adjustment instructions are supplied; 20kg capacity is only a verbal assertion without test evidence and is excluded from copy.

KV1 — Hero | Adjustable kitchen storage rack / Taobao (teaching example)
Task: identify the kitchen storage rack first; adjustment is secondary
Scene: complete rack and real adjustment relationship from structure references; preserve proportions and Style Lock
Copy: headline "Adjustable kitchen storage rack"; support "Adjust the width as needed"; no number tag
Layout: whole product first; upper-left product-name area and bottom annotation space
Text rendering method: Layout-reserved; designer adds typography in post-production
Copy-spec self-check: 4 headline words; one 5-word support line; zero number tags; KV1 category clear
Prompt: complete adjustable kitchen storage rack with its real adjustment relationship; supply the specified SKU drawing and preserve its shape/proportions; pale gray background, soft light, upper-left text space; no new structure, accessories, promotional text or watermark
Negative Prompt: distorted structure/logo, extra accessories, promotional text, watermark
Evidence/platform: drawing supports adjustment; 20kg needs proof and is excluded; current platform rules unverified

For cross-border work, preserve factual scope while adapting units, use cases, cultural context, language length and purchase reasoning rather than translating literally.

On-image copy and text rendering

Copy specifications (defaults, not platform rules)

Surface	Suggested ceiling
Main image	One headline (≤12 Chinese characters / ≤6 English words) + at most two support lines (each ≤18 characters / ≤10 words) + at most one number or evidence tag
Independent PDP module	One headline (≤12 characters / ≤6 words) + at most three points (each ≤18 characters / ≤12 words); number/evidence content counts toward the points
Disclaimers/annotations	Distinct readable bottom area; material conditions also sit beside the relevant claim

Count numbers and punctuation, exclude whitespace for Chinese; adapt other languages to comparable visual width. Number/evidence tags are optional and must already be in the evidence ledger. Keep specification tables and FAQs separate. Cut or split images/modules when over the limit; do not shrink type to force a fit. Explicit layout needs may adjust these suggestions with a reason, without relaxing factual or platform constraints.

For the copy-spec self-check, count each final headline and individual support/point string. Count English words by whitespace. With execution tools, run python scripts/count-copy.py --language en "Final headline" "Final support" (omit the language flag for Chinese), or use an equivalent counter; claiming a manual recount is insufficient. Without a counter, state that lengths await tool verification instead of estimating them. Recalculate after edits.

• At a 1080px canvas width, use at least 72px for headlines, 42px for support, and 36px for necessary annotations/disclaimers; scale proportionally. This is the skill's internal readability baseline, not a platform rule. Check at about 360px display width. Shorten copy, split modules or move parameters into page text rather than shrinking type.
• Put disclaimers/evidence annotations in a distinct readable bottom area with a safe inset and sufficient contrast. Material price/efficacy conditions must also sit next to the claim. Preserve required approved statements verbatim; these are exempt from marketing word limits, so adapt the layout.

Listing pages, first screens and covers

Where attention is scarcest, secure three things: a dominant subject, a first touchpoint that identifies the category at a glance, and key text clear of fixed areas the platform may cover (campaign badges, stickers, platform icons). Suggested values: product about 60% of the frame and still recognizable at about 200px; text concentrated in the top third, with roughly the bottom 20% left clear. Covered areas vary by platform and period; check the platform's current specification.

Text rendering method (choose exactly one per card)

• Layout-reserved (default): specify text-safe-zone position, proportion and background brightness; generate no new promotional text, prices, badges or watermarks. Deliver a separate final text layer for designer/layout-tool post-production. Preserve existing packaging text and logos, compositing the original product image if needed.
• Model-rendered (exception): only when the user explicitly asks and the text is very short (default ≤8 Chinese characters or one English headline of ≤6 words). Quote exact strings, language, type style, position and hierarchy. For longer text, deliver Layout-reserved and explain why.
• Include text acceptance and fallback in every card. Inspect every character after model rendering; typos, missing text, gibberish or packaging rewrites are Fail conditions. Fall back to Layout-reserved plus post-production typography. Without an inspected image, state that text rendering is pending acceptance. If the platform disallows overlays, provide no new promotional text.

Copy quality self-check (run on every line of copy)

1. Selling-point voice: a KV1 product name establishes identity and must not be deleted for lacking a sales pitch; ask of other marketing sentences "what is this to me, and why is it worth buying?"; cut or rewrite if not. Say why it is worth buying, not how it is constructed.
2. Information gain: apart from the first-image name and necessary specifications/conditions, text prioritizes supported information the image cannot say (motion, occasion, touch, specifications, pairing) and never restates the picture; headline and support lines do not repeat each other.
3. Numbers with perception: pair a number with position, bodily feel or a reference object; a headline is never a bare parameter. For fit content (size, specification, audience) give bands, a conclusion and the consequence, not "choose by your own situation". Use only data present in the materials for feel and thresholds; if absent, omit the claim or list the evidence gap outside the copy rather than invent or repeat the preference questionnaire.
4. Consumer language: jargon appears only in the specification area, and promotional copy uses words the buyer would use ("visual focus" and "information hierarchy" stay out of the image); read aloud and fix anything unnatural; cut a point rather than keep a fragment.
5. Question checks: every FAQ or self-ask question must satisfy all three: the evidence can answer it (if not, do not ask it, and do not patch with a disclaimer); it gives a clear judgment, not "depends on your situation"; and it addresses only concerns buyers of this category really have, never another category's.
6. Clarifications in place: clarifications about sets, pairings and specifications ("single item", "accessories not included") belong in a spec table or readable annotation area, not selling-point space; show what is actually included when showing pairings. Material conditions about included items or price remain beside the relevant picture/claim, never hidden in small print.

Details and category-neutral examples are in references/copy-quality-rules.md; category-specific notes are in references/category-notes.md. With an execution tool, run python scripts/scan-final-copy.py "final copy 1" "final copy 2" before delivery to scan for placeholders and internal notes.

Step 6 — Five-part quality gate

Complete this step before delivering the final draft. The fixed order is draft → self-review → rewrite → deliver, done inside the same reply; never show an under-threshold draft as final.

1. Draft the complete content. With an execution tool, run the mechanical checks first: python scripts/count-copy.py ... and python scripts/scan-final-copy.py ....
2. Check hard blockers, then verify each dimension's required conditions and score it.
3. If any dimension is below 80, rewrite only the content behind that dimension and score again; at most two rewrite rounds.
4. Once every dimension passes, deliver the final draft with the self-review record (format below).
5. If a dimension is still below 80 after two rounds, deliver the parts that pass, and state outside the copy which dimension failed, why, and what materials are needed; never pretend it passed or label under-threshold content final.
6. In Standard mode with staged delivery, self-review before each stage; mark dimensions that do not apply to that stage as "not applicable at this stage" with a reason.

Review the current delivery layer: a brief can verify naming, evidence, copy and layout specifications without establishing rendered-image quality. Unrendered images, packaging fidelity and actual mobile previews remain pending; self-scores never replace inspection. A naming-clarification reply does not require the full scoring record.

Hard blockers first:

• Invented or untraceable qualifications, tests, certification, sales, reviews, rankings, patents or numbers.
• Product identity conflicts with source materials.
• Needs proof written as fact.
• Unverified current platform rules stated as certain requirements.
• Unsupported medical/health guarantees or claims outside approved scope.
• Mainland China advertising with prohibited claims, missing statutory ad approval, or required citation sources/conditions hidden in internal notes or unreadable text.
• Final copy still contains placeholders or internal notes ("to be added", "to confirm", "(not provided)"), or clarifications sitting in selling-point space.
• Process, material or performance inferred from an image or not checked against the real item is written as fact, or kept despite conflicting with the item.
• An image prompt that depends on context or lacks product structure or forbidden changes and cannot be used on its own.
• No self-review record, or any dimension below 80 in the record while the draft is still delivered as final.

Five Pass/Fail dimensions and the 80-point line

Score each dimension 0–100 independently; 80 is the passing line. Scores are anchored to required conditions: only when every required condition below is met may a dimension score 80 or above; if any is unmet, the dimension is capped at 79 and must be rewritten. High scores elsewhere cannot compensate. Above 80, bonus items decide the score, and scores of 90+ need written evidence.

Dimension	Required conditions (all met for ≥80)	Bonus items (80–100)
Evidence & compliance	① every factual, numerical or qualification claim traces to the ledger ② Needs-proof content is not written as fact ③ China ad compliance check passed (when applicable) ④ no placeholders or internal notes in final copy ⑤ process, material and performance follow the materials' wording and were checked against the real item	sources cite page or image number; risky lines have substitutes
Conversion logic	① each image has its own decision task ② top reasons are clear (no padding to three) ③ naming is resolved in Step 0; permitted KV1 headlines name the product before secondary benefits, no-text images have a recognition basis, and later images need not repeat the name ④ copy passes the six-point quality self-check	support lines add what the picture cannot say; fit content gives bands and a conclusion
Visual hierarchy	① a clear first focus ② text within the copy-specification limits ③ reasonable subject proportion ④ the brief specifies mobile preview/type checks; inspect existing renders and mark unrendered images pending	reading order and emphasis of numbers/evidence are explained
Consistency	① Style Lock followed; no drift in style, palette or light ② product identity matches references ③ type and copy tone match across the set	every prompt repeats the Style Lock
Executability	① no guessing about essential elements ② every prompt is self-contained (structure, materials, lighting, composition, whitespace, forbidden changes) ③ every card states its text rendering method and copy-spec self-check ④ no "TBD" or "same as above"	targeted Negative Prompts; text acceptance and fallback stated

Check word limits, rendering responsibilities, disclaimer placement and mobile preview. The 80-point line is this skill's internal delivery threshold and must be anchored to the required conditions above, not an unanchored "80/100" feeling. If the user explicitly wants a numerical score to display, follow section 7 of references/output-contracts.md.

Self-review record (must accompany every delivery)

Append after the final draft. Without it, self-review is treated as not done, which is a hard blocker:

### Self-review record
| Dimension | Score | Evidence (one sentence: which required conditions are met) | Rewritten |
|---|---|---|---|
| Evidence & compliance | 85 | … | none |
| Conversion logic | 82 | … | rewrote KV3 headline |
| Visual hierarchy | 80 | … | none |
| Consistency | 84 | … | none |
| Executability | 83 | … | none |
- Mechanical checks: scan-final-copy passed / count-copy verified / no execution tool, not run
- Rewrite rounds: 0 / 1 / 2
- Items below threshold: none (or dimension, reason, materials needed)

Compliance and currency

Read references/compliance-rules.md for compliance work and locate official sources by category in references/source-notes.md. Distinguish reviewed provisions from login-only policy entry points. An accessible entry point does not verify a specific rule.

Mainland China main-image and PDP advertising rules

Before writing or revising main-image/PDP copy for mainland China on Taobao/Tmall, JD, Pinduoduo, Douyin Shop, Xiaohongshu, WeChat Channels/WeChat Shop or similar surfaces, read section 10 of references/compliance-rules.md and check the applicable PRC Advertising Law sources. Target market determines the rules regardless of output language.

• Establish the product category and permitted claims before converting materials into benefits. Do not copy unlawful or misleading source wording. Reference fidelity does not excuse problematic packaging claims: flag the blocker and obtain compliant assets without silently altering the product.

• Default China marketing copy excludes absolute superlatives such as “国家级、最高级、最佳”, unsupported leadership/uniqueness/lowest-price claims and medical-effect promises. Do not evade restrictions with homophones, pinyin, split characters, symbols or visual implications. For a potentially substantiated exception, provide a plain factual alternative and put the original outside final copy for human review. Human review cannot authorize a prohibited claim. Distinguish objective instructions such as “第一步” from superiority marketing; do not ban individual characters mechanically.

• Ledger the evidence for numbers, reports, rankings and comparisons, then check legal permissibility. Required citation sources, sample/conditions, scope and validity belong in consumer-visible text, not only the ledger. Distinguish advertiser self-tests from external citations; AI-generated data is not verified evidence.

• Prices, gifts, promotions and service promises must match confirmed terms. Keep material conditions beside the claim, not hidden in low-contrast small print or an “actual product prevails” disclaimer. Check patent/certification scope and endorsement/testimonial rights; invent no official approval.

• Ordinary goods cannot use disease-treatment claims or wording that confuses them with drugs/devices. If statutory ad approval is required but missing, provide a verification checklist and non-publishable structure only. Do not freely edit an approved advertisement; preserve applicable mandatory statements.

• Every card includes a China advertising compliance check, with a separate final text layer and risk/source notes. Blocked or pending-human-review claims are not publishable. Continue supported portions; image acceptance remains pending until inspected. Do not promise Advertising Law certification or platform approval.

• A disclaimer cannot repair an impermissible claim.

• Absolute, comparison, price, certification, health, efficacy and review claims depend on evidence and target market.

• Ordinary food, health food/supplements, medical devices, care products and children's goods require different claim boundaries.

• Laws, platform rules and category requirements change; prefer current official documents when available. This framework is not legal advice.

Reference loading

• references/compliance-rules.md — section 10 is mandatory before all mainland China main-image/PDP copy, including isolated revisions; other markets load relevant compliance/category sections as needed.
• references/platform-playbooks.md — platform and cross-border differences.
• references/output-contracts.md — mandatory before complete plans/strategies/storyboards or formal audits in every mode; also holds the kickoff questionnaire template.
• references/source-notes.md — applicable official laws/policies, price, health and children's claims.
• references/copy-quality-rules.md — check finished copy against the detailed rules and examples before delivery.
• references/category-notes.md — category-specific notes (currently jewelry).
• examples/README.md — example requests and expected behavior.

Key failure modes

• Over-questioning: questions continue beyond the kickoff questionnaire, or the same dimension is asked repeatedly. Proceed on the default assumptions.
• Skipping the questionnaire: platform, style, selling points or similar dimensions are clearly missing, yet neither a questionnaire nor labeled assumptions appear.
• Mechanical checkpoints: forcing stops despite one-shot intent. Follow the selected mode.
• Fixed five images: padding an image set without decision tasks. Adapt count.
• Selling-point overload: multiple purchase reasons per image. Use one primary task.
• Feature-only first image: mood or form replaces the product name. Identify the confirmed name/category in KV1 before secondary features; for no-text images, check the product itself is recognizable.
• Starting without a name: omitting the name or presenting a guess as an original before asking. Resolve user-supplied versus authorized image-based naming at the start and record its source.
• Outdated policy: presenting historical practice as current requirements. Verify official sources.
• Keyword-only compliance: deleting words without checking the underlying claim. Review fact -> evidence -> consumer interpretation.
• Literal localization: translating without adapting purchase reasoning, units and context.
• Reference drift: changing packaging/structure for aesthetics. Preserve real product identity.
• Example treated as proof: sample artwork does not substantiate its own claims. Check original materials.
• All text handed to the model: long text is error-prone. Choose one method per card; default to Layout-reserved, with model rendering only for explicitly requested short text followed by character-by-character review.
• Copy overload: headline, support, tags and disclaimers crowd one image. Cut or split according to the copy specifications, never shrink to fit.
• Promotion fabrication: no evidence means no invented reference price, lowest-price claim, time limit, inventory urgency, discount or gift.
• Evidence mistaken for legality: copying a documented claim or changing characters without removing a prohibited medical/guarantee meaning. Check category and legal restrictions first; make citation conditions visible.
• Producer's viewpoint: leading with construction, parameters and proof before why the buyer should care. Organize by the buyer's decision order.
• Working draft as deliverable: placeholders and internal notes printed on finished images. Scan before delivery and delete missing items entirely.
• Inference as fact: guessing process from the image, or copying merchant claims without checking the photo. Use the materials' wording; when it conflicts with the real item, the item prevails.
• Reference as decoration: reading the material but still writing around the product, learning only surface ornament. Distill hard constraints before writing.
• Ignoring downstream use: delivering concept descriptions instead of executable input. Prompts must be self-contained, complete and explicit about what is forbidden.
• Skipping self-review: delivering a final draft without a self-review record, or omitting it for "one-shot" or "no interruptions" requests. Self-review must be done before delivery.
• Rubber-stamp self-review: writing "all pass" without evidence or inventing score differences. State the required conditions met; equal evidence may receive equal scores. Never mark unseen renders accepted.

---

E-commerce Visual Copywriting v3.0 · Canonical runtime entrypoint: SKILL.md