331 SupercmoHQ

analyzing-products Skill

把产品规范化为可直接用于生成的信息——从电商 URL(干净的描述加精选产品图)或仅凭一张照片(品类、用途、可动/可开合部件及关键视觉细节)。当需要把产品 URL 或照片转换成图像/视频生成的输入,或其他技能在生成前需要产品信息时使用。

安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。

查看源码

技能指令原文(SKILL.md)

Product Analysis

Turn a product — an e-commerce URL or a bare photo — into clean, reusable facts for downstream
image/video generation: a tight description, curated product images, and how the product is physically
used. It's a building block: other skills call it at the input stage, before any generation.

Workflow

Step 1: Pick the mode and run it

| Input | Reference | What you produce |
| ----- | --------- | ---------------- |
| An e-commerce URL (Amazon, Shopify, AliExpress, any product page) | references/url-extract.md | A two-paragraph description + up to 5 downloaded, filtered product images |
| A product photo only (no URL, no description) | references/photo-analysis.md | Category + how it's used + moving/opening parts + key visual details |

Read only the matching reference and follow it end to end. If both a URL and a photo are given, run
the URL mode (richer) and keep the photo as one more reference image; if neither is given, there's
nothing to analyze — ask for one. Don't pause for confirmation — a URL (or photo) plus generation
intent means extract and proceed.

Step 2: Hand off

Return the result to whoever called you, ready to drop into generation:

  • URL mode → the description and the kept image files (local paths, usable as reference images).
  • Photo mode → the category, how it's used, any moving or opening parts, and the key visual details.

Don't rank the product's market position — the calling skill decides that from packaging cues. Your
job is the objective facts.

Edge cases

  • The URL can't be extracted (no result, or the extractor isn't set up) → ask for a product photo

instead and switch to photo mode.

  • Every image fails the filter (faces, wrong variant, not a product shot) → keep the single

cleanest, or hand off the description alone and say plainly that no clean image survived.

  • A supplied photo is too unclear to read (blurry, cropped, ambiguous) → say what you can't

determine and ask for a clearer shot rather than guessing the mechanic.

  • Neither a URL nor a photo → ask for one; there is nothing to analyze.

Reference

  • references/url-extract.md — the URL pipeline: url_extraction → download the images → filter them

with image_analysis → write the description.

  • references/photo-analysis.md — the photo pipeline: category, how it's used, moving/opening parts,

and the visual details to preserve.