Skip to content
Back to search

ec-visual

SkillWorks with Claude CodeGitHub

"Use when the user uploads product images and visual references to create or improve marketplace head images and product detail pages. Includes consumer-angle mapping, fixed-screen detail-page expression planning, visual proof selection, brand/product-adapted color, prompts, and optional final image generation."

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/oldred-byte--ec-visual-skill/badge.svg)](/listings/oldred-byte--ec-visual-skill)

HTML

<a href="/listings/oldred-byte--ec-visual-skill"><img src="/api/listings/oldred-byte--ec-visual-skill/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 102 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. ---
  2. name: ec-visual
  3. description: "Use when the user uploads product images and visual references to create or improve marketplace head images and product detail pages. Includes consumer-angle mapping, fixed-screen detail-page expression planning, visual proof selection, brand/product-adapted color, prompts, and optional final image generation."
  4. ---
  5. # EC Visual
  6. ## Overview
  7. Create e-commerce visual systems from product image(s), reference image(s), and optional product facts: 5 marketplace head images and/or 10 vertical detail-page screens. References are used as visual language sources, not rigid templates. Product facts, any user-approved screen directions, and each screen's consumer-facing expression task decide the content; reference DNA, product/brand color, and proof grammar decide how the set stays coherent and commercially useful.
  8. ## When To Use
  9. Use this skill when the user asks to:
  10. - Generate product head images, Taobao/Tmall/JD/Douyin main images, product main images, or marketplace listing images.
  11. - Generate detail-page screens or a complete set containing both 5 head images and 10 detail-page screens.
  12. - Turn product images plus poster/head-image/detail-page references into planned image prompts or final images.
  13. - Build reusable e-commerce visual prompts from product facts, documents, and visual references.
  14. If the user asks for only one e-commerce image, still use the relevant input, reference, color, product-use, and visual-proof rules.
  15. ## Workflow
  16. 1. Inspect all product images, reference images, and uploaded documents. Classify image roles with `references/reference-analysis.md`.
  17. 2. Route product information with `references/input-routing.md`: if the user uploaded only images and no product facts, ask once whether they have selling points, product information, brand colors, parameters, or documents to provide before planning. Complete facts are followed, partial facts are preserved and improved, missing facts are inferred conservatively from product image and category context.
  18. 3. Capture any user-provided page structure or screen directions as hard scaffolding. Do not replace an approved 10-screen direction with a new generic sequence; translate each fixed screen direction into a stronger consumer-facing expression.
  19. 4. Build a consumer-angle map with `references/input-routing.md`: category anxieties, desired outcomes, body/use moments, objections, and what evidence would make a buyer believe each point.
  20. 5. Extract a set-level visual DNA contract from the strongest reference(s) using `references/visual-dna.md`. Treat the reference as visual language only (light behavior, typography attitude, density rhythm, commercial temperature). Reference scene/setting elements (specific fabric, specific surface, specific location, specific background texture) are NOT inherited by default and may appear in at most 30% of outputs.
  21. 6. Build the color system with `references/color-adaptation.md`: user-provided brand colors first when supplied, otherwise product identity colors first. Reference colors are borrowed only as organization logic unless they fit the product/brand.
  22. 7. Build product-use plausibility, interaction translation, claim safety, and model consistency plans with `references/reference-analysis.md`.
  23. 8. For every image/screen, write a Picture Solo Statement first using `references/visual-proof-grammar.md`: one sentence describing the concrete picture, followed by what message it proves when all copy is covered. If the picture cannot convey the message alone, redesign the picture before continuing.
  24. 9. Decide product presence per screen using `references/message-picture-patterns.md`: match the screen's message to a pattern, then follow that pattern's product-exposure rule. The product does NOT appear by default; it appears only when the message requires it.
  25. 10. **Unified Plan + Approval Gate (mandatory)**: Plan head images and detail pages TOGETHER as one batch. Output ONE unified approval table that covers all 5 head images AND all 10 detail-page screens — each row containing: picture solo statement, pattern match, main subject, product exposure, scene setting, AND draft consumer-voice copy. User reviews this single table and approves or modifies before ANY prompt is written or ANY image is generated. Do not split head-image approval from detail-page approval. See `references/head-images.md` and `references/detail-pages.md` for table structure. Skip only when the user explicitly says to skip review.
  26. 11. For final image deliverables, write one prompt per image/screen and use the image generation tool. Follow `references/production-loop.md` for generation, validation, regeneration, saving, reporting, and the post-iteration skill-learning hook.
  27. ## Direct Generation Contract
  28. When the user asks to make actual image deliverables:
  29. - **Create a task folder first** before any generation. Default location: `~/Desktop/`. Folder name format: `YYYYMMDD-HHMM-<产品名或类别>` (e.g., `20260705-1730-英式辅食`). All outputs for this task go ONLY into this folder. Never write into a folder containing previous-task files. See `references/production-loop.md` for full Task Folder Isolation rules.
  30. - Generate each requested deliverable as its own independent model-native image.
  31. - Save head images as `<task-folder>/主图/01-*.png` through `<task-folder>/主图/05-*.png`.
  32. - Save detail pages as `<task-folder>/详情页/01-*.png` through `<task-folder>/详情页/10-*.png`.
  33. - Save contact sheets to `<task-folder>/记录/`, validation report to `<task-folder>/验证报告.md`.
  34. - Use contact sheets only as review artifacts after independent final files exist.
  35. - Do not use deterministic local compositing, local text overlay, local retouching, or local layout repair on final deliverables.
  36. - Validate each file and the full set before reporting completion.
  37. ## Non-Negotiable Rules
  38. - Product accuracy outranks reference style.
  39. - **Task folder isolation (hard rule)**: every new task creates a fresh folder at `~/Desktop/YYYYMMDD-HHMM-<产品名或类别>/`. All outputs (主图, 详情页, 记录, 验证报告) live inside this folder. Never write into a folder containing files from a previous task. See `references/production-loop.md`.
  40. - **Copy is consumer-facing, not explanatory (hard rule)**: every line of copy (headline, sub-headline, support copy, label) is written from the consumer's perspective — what they feel, get, or stop worrying about. Never write manufacturer/explanatory/spec-sheet voice ("采用XX技术", "本产品具有XX功能", "符合XX标准", "XXX工艺打造"). See `references/visual-proof-grammar.md` Copy Voice section.
  41. - The visual must prove the message; do not add a model, scene, chart, macro, label, or decorative module unless it helps the viewer understand or believe the core message.
  42. - **Picture Solo Test (hard gate)**: every output must work with all copy covered. Before writing a prompt, write one sentence: "The picture is [concrete scene], and with all text hidden it still shows [message]." If you cannot, the picture is failing its job — redesign it before prompting.
  43. - **Product presence is decided by message, not by default**: the product/packaging does NOT appear in every output. It appears only when the message requires the packaging to be visible. Multi-flavor, multi-ingredient, micro-material, mechanism, sensory, scene-fit, and lifestyle-proof screens usually have NO packaging in the main visual. Use `references/message-picture-patterns.md` to decide.
  44. - **No icons ever**: never include icon-shaped elements that "express a selling point" (checkmarks, gear icons, lightning, shield icons, ingredient bubbles, feature pictograms, decorative arc systems). Icons added later by the user are out of skill scope. A screen is composed of: main visual + main copy + optional short support copy. It is NOT composed of: icons, icon rows, icon grids, icon bubbles, badge ribbons, or pictogram systems.
  45. - **Reference is language, not scene**: inherit light behavior, typography attitude, density rhythm, commercial temperature, and module rhythm from the reference. Do NOT inherit specific scene elements (fabric type, surface texture, location, background pattern). Reference scene elements may appear in at most 30% of outputs across the set; the remaining outputs must use a different main visual.
  46. - **Unified approval gate (hard rule)**: head images and detail pages are planned and reviewed TOGETHER. Output ONE approval table covering all 5 + 10 outputs (with draft consumer-voice copy) before any prompt is written. User approves the whole batch in one pass. Do not generate head images first and detail pages separately without a unified approval.
  47. - **Set-level diversity is mandatory**: 5 head images must use ≥4 distinct main visuals; 10 detail-page screens must use ≥7 distinct main visuals. "Main visual" = scene + main subject + composition family. Same main visual repeated = template failure, regenerate.
  48. - The product image is a source of truth, not a requirement that the whole product must dominate every screen. Some screens may be led by skin, hand interaction, body contact, macro structure, mechanism, result state, scene context, or a partial product crop when that better proves the screen's task.
  49. - References provide visual DNA, layout logic, proof methods, rhythm, typography attitude, and mood; they do not provide copied brands, people, claims, prices, logos, certifications, product-specific actions, OR specific scene/setting elements.
  50. - Detail pages are content-led and sequence-led. A reference may shape style and consistency, but each screen's message decides its composition. Detail pages are NOT vertical extensions of head images; they have their own sequence logic.
  51. - If the user has already defined the detail-page screen directions, preserve those directions and improve the expression inside each screen instead of inventing a new structure.
  52. - Every human-product interaction must be physically and behaviorally plausible for the actual product category. Translate reference action intent instead of copying poses.
  53. - Color is rebuilt around supplied brand colors when present, otherwise around the product identity and category. Do not copy reference dominant colors by default.
  54. - Avoid mechanical repetition. A coherent set should share visual DNA while varying proof method, composition, density, and product-human relationship by role.
  55. - Do not invent prices, official status, certifications, specs, ratings, medical effects, or unsupported performance claims.
  56. ## Output Shape
  57. Default deliverable when not directly generating:
  58. 1. Input and reference role analysis.
  59. 2. Product information route and claim-safety summary.
  60. 3. Consumer-angle map and screen-structure capture.
  61. 4. Set-level visual DNA and color system.
  62. 5. **Unified approval table**: 5 head images + 10 detail-page screens, each with picture solo statement, pattern match, main subject, product exposure, scene setting, and draft consumer-voice copy.
  63. 6. Generation prompts for requested outputs.
  64. Default deliverable when directly generating:
  65. 1. **Output the unified approval table** (5 head images + 10 detail-page screens with draft copy) and ask for user approval before any image generation. Skip only if the user explicitly says to skip review.
  66. 2. Brief planning summary after approval or when review is skipped.
  67. 3. Generate and save requested independent final image files into the task folder.
  68. 4. Create contact sheets only as secondary review artifacts.
  69. 5. Report saved paths and validation/regeneration notes.
  70. ## References
  71. Read only the needed reference files:
  72. - `references/input-routing.md`: product facts, documents, completeness routes, and selling-point extraction.
  73. - `references/reference-analysis.md`: image role classification, reference transfer, product-use plausibility, model consistency, claim safety, and the reference-is-language-not-scene rule.
  74. - `references/visual-dna.md`: set-level consistency, reference weighting, per-screen DNA inheritance, and set-level main-visual diversity thresholds.
  75. - `references/color-adaptation.md`: brand/product-first color system and anti-color-copy rules.
  76. - `references/visual-proof-grammar.md`: Picture Solo Test, message-first visual proof, product-presence decision, no-icon rule, consumer expression, and density rhythm.
  77. - `references/message-picture-patterns.md`: high-frequency message-to-picture patterns (multi-flavor, ingredients, sensory, mechanism, scene, trust, etc.) with concrete picture patterns and product-presence rules.
  78. - `references/head-images.md`: 5-head-image structure, 5-image main-visual diversity table, and prompt requirements.
  79. - `references/detail-pages.md`: fixed 10-screen/detail-page expression planning, 10-screen main-visual diversity table, approval gate, and prompt requirements.
  80. - `references/production-loop.md`: autonomous generation, validation, iteration, output hygiene, set-level pre-check, and skill feedback loop.
README.md · README · 133 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. # EC Visual Skill 🛒✨
  2. 让 Codex 帮你把「产品图 + 参考图」整理成一套完整的电商视觉生产流程。
  3. 它适合用来做商品主图、详情页分屏、参考图拆解、生成提示词,以及成套图片的检查和返工。你给它产品和参考,它会像一个耐心的电商视觉导演一样,先看清楚产品,再拆参考,再安排每一张图该承担什么销售任务。
  4. ## ✨ 它能做什么
  5. - 规划 **5 张电商主图**
  6. - 规划 **10 屏竖版详情页**
  7. - 分析参考图的布局、镜头、文字节奏和商业气质
  8. - 根据产品重新适配颜色和视觉系统
  9. - 为每张图写独立的生成 prompt
  10. - 生成后检查产品一致性、文字可读性、风格稳定性
  11. - 对失败图片进行 prompt 修正和重新生成
  12. - 保存最终图片和验证记录
  13. ## 🧭 它怎么工作
  14. 1. **读取素材**
  15. 识别产品图、参考图、产品事实和目标用途。
  16. 2. **拆解参考图(语言,不是场景)**
  17. 只提取参考图的视觉语言——光感、字体态度、密度节奏、商业温度、模块节奏。**不复制参考图的具体场景元素**(特定布料、场地、纹理)。
  18. 3. **建立产品规则**
  19. 产品外观、颜色、材质、结构、使用方式优先保持准确。颜色以产品/品牌为准,不照搬参考图主色。
  20. 4. **统一规划 5 张主图 + 10 屏详情页**
  21. 主图和详情页一起规划、一起审核。每一张/每一屏都先匹配 message-to-picture pattern,再写 Picture Solo Statement(画面独行声明),最后写消费者视角的草稿文案。
  22. 5. **统一审核关卡**
  23. 输出**一张统一审核表**,包含 5 张主图 + 10 屏详情页的全部信息(画面独行声明、pattern、产品曝光、草稿文案)。用户审核通过后才开始生成。
  24. 6. **逐张写 prompt**
  25. 一张图一条 prompt,包含画面独行声明、pattern、产品曝光规则、消费者视角文案锁、反 icon 声明。
  26. 7. **生成、检查、修正**
  27. 检查项目包括:画面独行测试、消费者视角文案、无 icon、产品曝光匹配、场景模板化、集级多样性。
  28. 8. **隔离保存**
  29. 每个任务自动创建独立文件夹 `~/Desktop/YYYYMMDD-HHMM-<产品名>/`,主图、详情页、记录、验证报告都进去,杜绝混批。
  30. ## 🔒 v2 核心硬规则
  31. 这一版针对实际使用中暴露的问题做了 8 条硬规则升级:
  32. - **Picture Solo Test(画面独行测试)**:每张图必须能盖住所有文案,画面独自传达 message。盖不住就重做画面,不许用文案补。
  33. - **Product-Presence Rule(产品按需出现)**:包装默认不出现。多口味、天然成分、微观材质、机制、感官这些 message 都不该有包装。
  34. - **No-Icon Rule(绝对禁 icon)**:不允许任何 icon 形态——勾选、齿轮、闪电、盾牌、成分泡、特征图标行、装饰弧线。icon 用户后面自己加。
  35. - **Reference Is Language, Not Scene(参考是语言不是场景)**:参考图的场景元素(野餐布、草地、大理石)最多只能在 30% 的图里出现。
  36. - **Set-Level Diversity(集级多样性)**:5 张主图至少 4 种不同主视觉;10 屏详情页至少 7 种。同构直接重做。
  37. - **Copy Voice(消费者视角文案)**:所有文案从消费者视角写,通过 "so what?" 测试。不许出现"采用XX技术"、"本产品具有XX功能"、"符合XX标准"这种甲方口吻。
  38. - **Task Folder Isolation(任务文件夹隔离)**:每个新任务在桌面建独立文件夹,时间戳+产品名命名,整套输出都进去。
  39. - **Unified Approval Gate(统一审核关卡)**:主图+详情页一起出审核表,含草稿文案,审完才生成。
  40. ## 📦 默认输出
  41. 如果只需要方案,它会输出:
  42. - 输入与参考角色分析
  43. - 产品信息路由 + claim 安全小结
  44. - 消费者视角地图
  45. - 集级视觉 DNA + 配色系统
  46. - **统一审核表**(5 主图 + 10 详情页,含草稿文案)
  47. - 15 条生成 prompt
  48. 如果需要直接出图,它会继续完成:
  49. - 在桌面创建隔离任务文件夹
  50. - 逐张调用图片生成
  51. - 逐张检查(含画面独行、消费者文案、无 icon、产品曝光、场景模板)
  52. - 对不合格图片返工
  53. - 保存独立图片文件到任务文件夹
  54. - 生成验证说明
  55. ## 🧱 仓库结构
  56. ```text
  57. .
  58. ├── SKILL.md
  59. ├── agents/
  60. │ └── openai.yaml
  61. └── references/
  62. ├── color-adaptation.md # 配色:以产品/品牌为准
  63. ├── detail-pages.md # 10 屏详情页 + 多样性看板
  64. ├── head-images.md # 5 张主图 + 多样性看板
  65. ├── input-routing.md # 产品事实路由 + 消费者地图
  66. ├── message-picture-patterns.md # 12 个高频 message → 画面 pattern 库
  67. ├── production-loop.md # 生成循环 + 任务文件夹 + 验证
  68. ├── reference-analysis.md # 参考图角色 + flavor contract
  69. ├── visual-dna.md # 参考是语言不是场景 + 多样性阈值
  70. └── visual-proof-grammar.md # Picture Solo + 产品按需 + 无 icon + 文案视角
  71. ```
  72. ## ⚠️ 使用边界
  73. 这个 skill 会尽量保护商业视觉的准确性和安全性:
  74. - 不编造价格、认证、检测结果、医疗效果
  75. - 不复制参考图的品牌、原文案、人物和具体承诺
  76. - 不为了追参考风格牺牲产品准确性
  77. - 不把 contact sheet 当成最终交付图
  78. - 不用本地贴字或拼图修补最终商业图
  79. - 不复制参考图的具体场景元素(野餐布、大理石等)超过全集 30%
  80. - 不在画面里放任何 icon 形态
  81. - 不写"采用XX技术"、"本产品具有XX功能"这种解释型文案
  82. - 不在没审核的情况下直接生成图
  83. ## 📜 Changelog
  84. - **v2 (2026-07)** — 加入 8 条硬规则:Picture Solo Test、产品按需出现、绝对禁 icon、参考是语言不是场景、集级多样性、消费者视角文案、任务文件夹隔离、统一审核关卡。新增 4 个 reference 文件(input-routing、message-picture-patterns、visual-dna、visual-proof-grammar)。
  85. - **v1 (2026-05)** — 初版:基础 5 主图 + 10 详情页流程,含 reference 解析、配色适配、生产循环。
  86. ## 🚀 安装方式
  87. 复制到 Codex skills 目录:
  88. ```bash
  89. cp -R ec-visual ~/.agents/skills/ec-visual
  90. ```
  91. 如果你是直接 clone 这个仓库,也可以用软链接:
  92. ```bash
  93. ln -s "$(pwd)" ~/.agents/skills/ec-visual
  94. ```
  95. 然后重启 Codex,或按你的环境重新加载 skills。
  96. ## 🪪 License
  97. MIT

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.