// the find
feichanggege/ecommerce-visual-copywriting-skill
电商视觉文案设计SOP Skill - 让电商设计小白也能一键做出5年资深设计经验的商品图片.兼容 Claude / Codex / Cursor / Workbuddy 等 SKILL.md Agent. AI-powered e-commerce visual strategy skill for main images, PDP/A+, copywriting, localization, evidence-aware compliance and image prompts.
A SKILL.md prompt-engineering package that gives Claude/Codex/Cursor-style agents a structured methodology for producing e-commerce product photography briefs, image-generation prompts, and copywriting for platforms like Amazon, Shopify, and Taobao. It's aimed at sellers and designers who want an AI agent to turn raw product data into a storyboard of shot-by-shot creative direction, not at developers looking for a code library.
The Evidence Ledger + Reference Fidelity + Negative Prompt structure directly targets real failure modes of naive AI-generated product images: fabricated certifications, packaging/logo drift between reference and output, and unverifiable claims slipping into ad copy. The platform playbooks split 'stable visual strategy' from 'current platform rules,' explicitly telling the agent to verify size/compliance rules against live sources rather than trusting stale training data — a detail most prompt packs skip. It covers both domestic Chinese platforms (Taobao, JD, Pinduoduo, Douyin) and cross-border (Amazon, Shopify, TikTok Shop, Temu, Shopee, Lazada) with localization guidance that goes beyond translation (units, purchase logic, information density).
Despite the Python language tag and 811 stars, there's no actual software here — tools/verify-skill.py is just a structural/privacy linter for the markdown package, not a functional pipeline; all the 'safety boundaries' and compliance rules are instructions to an LLM with zero enforcement mechanism, so adherence depends entirely on the model's willingness to follow them, not on any code-level guardrail. Primary documentation is Chinese with English as a secondary companion file, so non-Chinese-speaking teams get a second-class experience. There's no way to evaluate output quality objectively since results are entirely a function of which model interprets the skill and how — the repo itself contains no evals, no test cases with expected outputs, just a JSON file of test prompts.