Skip to content
Back to search

huatu-skill

SkillWorks with Claude CodeGitHub

Turn articles, research, notes, data, or existing drafts into clear Xiaohongshu carousel images or self-contained ELI5 visual explainer HTML. Use for 小红书图文、小红书轮播图、信息卡片、知识型笔记、研究型长图, responsive mechanism explainers, local PNG/HTML rendering, and checks for typography, hierarchy, overflow, sources, simplification boundaries, and safe self-contained delivery.

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/ythx-101--huatu-skill/badge.svg)](/listings/ythx-101--huatu-skill)

HTML

<a href="/listings/ythx-101--huatu-skill"><img src="/api/listings/ythx-101--huatu-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 · 229 lines
  1. ---
  2. name: huatu-skill
  3. description: Turn articles, research, notes, data, or existing drafts into clear Xiaohongshu carousel images or self-contained ELI5 visual explainer HTML. Use for 小红书图文、小红书轮播图、信息卡片、知识型笔记、研究型长图, responsive mechanism explainers, local PNG/HTML rendering, and checks for typography, hierarchy, overflow, sources, simplification boundaries, and safe self-contained delivery.
  4. ---
  5. # 画图 Skill
  6. Build a carousel that helps the reader understand the meaning — before it looks coherent, before it passes the rules. A deck succeeds when the layout itself makes a relationship, mechanism, or judgment easier to see than the text alone would. Treat model interpretation, information architecture, visual hierarchy, deterministic rendering, and visual QA as one workflow. Constraints are guardrails, not goals: they prevent accidents, but a content-specific visual thesis creates meaning. When a rule and the reader's comprehension conflict, the rule yields.
  7. ## Deliver a product, not a demo
  8. Huatu has two product modes. In **carousel mode**, the final product is the rendered PNG set; HTML is the deterministic preview layer and JSON is the editable source. In **ELI5 HTML mode**, one self-contained responsive HTML file is the final product; browser screenshots and QA JSON are evidence. Neither mode may substitute a preflight result for the target artifact. Read [references/delivery-standard.md](references/delivery-standard.md) before creating or repairing either product.
  9. Use state words precisely: `concept`, `draft`, `candidate`, `release`, or `blocked`. Never call work complete when browser rendering, target-artifact inspection, or blocking-finding resolution is still pending. `--check-only` proves static/schema validity only. A render failure is `blocked`, not a license to hand off an unverified artifact.
  10. ## Choose the task path
  11. - Choose **ELI5 HTML mode** when the user wants one responsive visual mechanism explainer rather than numbered carousel images. Read and follow [references/eli5-html-mode.md](references/eli5-html-mode.md), then use `assets/eli5-explainer-starter.html` and `scripts/check_explainer.py`. Do not apply the carousel JSON/render workflow to this mode.
  12. - Otherwise choose **carousel mode** and continue below. For 知识型、机制型或研究型内容,默认选择 `mixed`(图文交替),并同时阅读 [references/image-led-design.md](references/image-led-design.md) 与 [references/diagram-system.md](references/diagram-system.md);只有用户显式要求全文字时才选择 `editorial`。
  13. - For source material or a long article, extract the thesis and build a slide-by-slide narrative.
  14. - For an existing outline or JSON spec, preserve the argument and improve page roles, density, and components.
  15. - For existing images, inspect every image first. Diagnose hierarchy, spacing, contrast, consistency, clipping, and sequence before rebuilding.
  16. - For a critique-only request, stop after the annotated diagnosis and prioritized fixes. Do not render unless requested.
  17. Read [references/content-blueprints.md](references/content-blueprints.md) when deciding the story sequence. For reference-driven or visually ambitious work, read [references/model-interpretation.md](references/model-interpretation.md) and write the design manifest before styling. When photographs, illustrations, collage, expressive objects, or material texture could carry meaning, also read [references/image-led-design.md](references/image-led-design.md). Read [references/kami-design.md](references/kami-design.md) (默认主题 v2.0 法则与 tokens), [references/design-contract.md](references/design-contract.md), [references/layout-system.md](references/layout-system.md), [references/visual-qa-rubric.md](references/visual-qa-rubric.md), and [references/delivery-standard.md](references/delivery-standard.md) before choosing styles or doing visual QA. Read [references/spec-format.md](references/spec-format.md) before writing a render spec.
  18. ## ELI5 HTML path
  19. For this mode, create 3–7 semantic scenes with one concept and one dominant inline SVG/figure per scene. Keep copy short, name facts and sources, and include an explicit simplification boundary. The file must use inline CSS/SVG only: no JavaScript, remote dependency, analytics, request-capable resource, credential, private identifier, host secret, or QR code.
  20. Run static preflight, then real Chromium validation:
  21. ```bash
  22. python3 <skill-dir>/scripts/check_explainer.py explainer.html --check-only
  23. python3 <skill-dir>/scripts/check_explainer.py explainer.html --output-dir explainer-qa
  24. ```
  25. The browser run is mandatory and must pass at 390×844 and 1365×768, producing `qa.json`, `preview-mobile.png`, and `preview-desktop.png` with no unsafe request attempt, horizontal overflow, browser error, or visible text below 12px. Inspect the HTML and both full-page screenshots after the last change. Deliver HTML as the product and QA/screenshots as evidence. Moshi/iPhone preview and release approval remain parent/human-owned; never start a public or persistent preview server.
  26. The numbered workflow below is the carousel path.
  27. ## 1. Establish the content contract
  28. Identify:
  29. - audience and desired reader action;
  30. - one-sentence thesis;
  31. - facts, quotes, and sources that may be presented as verified;
  32. - desired slide count, defaulting to 6–10 for a dense educational carousel;
  33. - author/footer text and any provided brand constraints.
  34. Infer these from supplied material when safe. Ask the human only when a missing choice would materially change the result. Never invent a statistic, quotation, source, credential, or platform rule to fill a layout.
  35. ## 2. Write the narrative before styling
  36. Create a storyboard with exactly one sentence per slide describing that slide's job. Give each slide one dominant message and select only the components needed to prove it.
  37. Prefer this rhythm for research-style content:
  38. 1. promise or tension;
  39. 2. evidence of relevance;
  40. 3. distinction or comparison;
  41. 4. mechanism or framework;
  42. 5. value, consequence, or case data;
  43. 6. disagreement or limitation;
  44. 7. the author's judgment and decision threshold;
  45. 8. deeper implication or local context;
  46. 9. conclusion and reader-specific action.
  47. Shorten or merge roles when fewer pages are appropriate. Do not force nine pages when the material cannot support them.
  48. ## 3. Interpret the content visually
  49. For reference-driven work, set `referenceDriven: true` and create a concise `designManifest`. The manifest records:
  50. - one visual thesis and intended emotional progression;
  51. - a semantic metaphor or a reason to use none;
  52. - one or two meaningful motifs;
  53. - deck rhythm, contrast plan, and type strategy;
  54. - one composition intent per slide;
  55. - recent habits and visual clichés to avoid;
  56. - why this direction fits this specific content.
  57. This is an inspectable design brief, not hidden reasoning. Use it to make decisions and later judge whether the render delivered the intended meaning.
  58. Choose a `visualMode` before choosing blocks:
  59. - `editorial` when language and evidence should remain primary;
  60. - `mixed` when image and structure should alternate;
  61. - `image-led`, `photo-diary`, `object-study`, or `poetic-poster` when a visual anchor should carry the page before explanatory text.
  62. 知识型、机制型和研究型内容默认使用 `mixed`;只有用户显式选择 `editorial` 时才改为全文字路径。任何非 `editorial` 模式的 manifest **必须**包含简短的 `materialSystem`、`imageCadence` 与逐页 `visualAnchor` 计划,不能只写一个模式名称。Do not satisfy this step by placing generic AI decoration behind existing cards. Each image needs a semantic role: hero, evidence, atmosphere, texture, motif, or transition.
  63. Before a full deck, create three two-page directions that differ in interpretation and composition—not merely colors. Prefer metaphor-led, typographic/editorial, and diagram/spatial directions when appropriate. Compare them with:
  64. ```bash
  65. python3 <skill-dir>/scripts/compare_directions.py direction-a.json direction-b.json direction-c.json
  66. ```
  67. If comparison reports `distinct: false`, revise the concepts or compositions before asking the human to choose a direction. As soon as a direction is selected, perform a real browser render of its two-page candidate. This is the runtime smoke test: discover missing Playwright/Chromium or broken local assets before writing the entire deck. If the smoke render cannot run, mark the release path `blocked` and report the exact prerequisite.
  68. ## 4. Apply a bounded visual system
  69. Use a 1080×1350 canvas unless the user specifies another target. For a single 3:4 cover (e.g. Xiaohongshu 封面), set the top-level `canvas` field to `{"width": 1080, "height": 1440}` and render a one-slide poster spec. **Default theme is Kami (v2.0)**: warm parchment canvas, ink-blue accent, warm-gray text scale, serif-led hierarchy (weight 500, no pseudo-bold), hairline borders, whisper-only shadows, no default italics, no yellow marker wash. Full rules, tokens, and anti-patterns live in `references/kami-design.md`. The old brick-red / pale-yellow default is retired.
  70. When to keep Kami: the default — especially when the user wants 高级感 / 文档感 / 克制 / 正式 / 专业. When to override the theme explicitly: the user asks for high-saturation, playful, or "小红书爆款" visual energy; then provide a custom `theme` in the spec, never by silently reverting the defaults.
  71. Keep these constraints as guardrails, not a fixed template:
  72. - one dominant headline per slide;
  73. - two to four content modules per slide;
  74. - no more than three highlight phrases per slide unless a table requires emphasis;
  75. - source labels smaller than body copy but still legible;
  76. - generous bottom clearance for platform overlays and pagination;
  77. - alternating dense and light pages across the sequence.
  78. For mixed or image-led decks:
  79. - begin with the image-text relationship, not a component quota;
  80. - avoid three consecutive pages that all require reading before seeing;
  81. - include a low-text visual pause when the narrative can support one;
  82. - limit the material vocabulary to two or three coherent treatments;
  83. - preserve asymmetry, cropping, and negative space when they improve the idea;
  84. - use original, user-provided, licensed, or public-domain assets only, and record provenance outside the slide when needed.
  85. Choose composition primitives because they express a relationship. Available archetypes are starting points (`editorial`, `poster`, `split`, `diagram`, `timeline`, `custom`); primitives may be combined within safe bounds. A custom composition must explain why breaking the normal grid helps the content.
  86. Abstract visual principles from references. Do not copy another creator's text, watermark, avatar, account name, signature, or distinctive identity treatment.
  87. ## 5. Create the render spec
  88. Copy [assets/starter-carousel.json](assets/starter-carousel.json) into the working output directory and replace its original example content. Keep user material outside the Skill directory.
  89. Use only the supported block types documented in [references/spec-format.md](references/spec-format.md), including the local-only `image` block. Treat all content as plain text; never insert user text into the HTML template as executable markup. Image sources must be local PNG, JPEG, WebP, or SVG paths relative to the spec or absolute local paths; remote URLs are rejected.
  90. For `mixed` or `image-led` specs, these are hard requirements:
  91. - 每张图先写一句关系命题:这张图让读者比读段落更快看见的流程、对比、趋势或分层是什么([references/diagram-system.md](references/diagram-system.md) 的一句话测试)。过不了测试就不画;
  92. - 知识型 deck 优先至少 2 张图卡(diagram 计入),但以一句话测试为准:内容只撑得住 0–1 张真有教学增量的图时,宁缺毋滥,不要为凑配额补装饰图;
  93. - 每张图必须有语义明确的 `alt` 与 schema 合法的 `role`,并在规划中承担 hero、evidence、annotate、contrast、echo 或 transition 之一,不能只负责装饰;
  94. - [assets/diagrams/](assets/diagrams/) 模板是候选不是默认:先按内容关系决定结构,再从模板里选最接近的作起点改写;内容被迫适配模板节点数时,删除、合并或重排节点,保留关系本身。
  95. Run structural validation before rendering:
  96. ```bash
  97. python3 <skill-dir>/scripts/render_carousel.py carousel.json --output-dir rendered --check-only
  98. ```
  99. Fix every reported spec error before continuing.
  100. ## 6. Render deterministically
  101. Render the deck with the bundled HTML/CSS template:
  102. ```bash
  103. python3 <skill-dir>/scripts/render_carousel.py carousel.json --output-dir rendered
  104. ```
  105. The renderer writes:
  106. - `carousel.html`, the editable local preview;
  107. - `slide-01.png`, `slide-02.png`, and so on;
  108. - `qa.json`, including dimensions and overflow results.
  109. - `design-manifest.json` for interpreted decks.
  110. For interpreted decks, it also compares a non-sensitive visual fingerprint with the most recent five entries in `design-history.jsonl` beside the spec. This produces a review warning rather than prohibiting justified reuse. Pass `--no-history` only for fixtures or reproducible tests.
  111. If Playwright or a Chromium browser is unavailable, report the exact missing prerequisite and mark the release `blocked`. Do not silently replace the renderer with AI-generated text images, and do not present HTML/JSON as the finished image product.
  112. ## 7. Perform visual QA
  113. Read `qa.json`, then inspect the rendered PNGs visually. For a deck of ten pages or fewer, inspect every page. Review the PNGs before reading the manifest or author rationale. At minimum verify:
  114. Comprehension probes come first — each is blocking, with the same weight as overflow:
  115. - whether the slide still carries mood or meaning when its explanatory copy is covered (遮文仍懂);
  116. - whether every image and diagram adds a teaching increment: it makes a flow, comparison, trend, or layer visible faster than the paragraph would (教学增量,diagram 一句话测试);
  117. - whether every image card has a visual anchor and first reading target visible within one second at 1× phone size.
  118. Then structure and craft:
  119. - title hierarchy and line breaks;
  120. - no clipped or overlapping text;
  121. - readable source labels and table cells;
  122. - consistent margins, rules, card radii, and page numbers;
  123. - restrained highlighting and accent usage;
  124. - sequence-level density and rhythm;
  125. - accurate author/footer text;
  126. - absence of copied identity elements;
  127. - semantic role, focal point, crop, and provenance of every image;
  128. - whether the image adds a relationship or proof instead of repeating nearby text;
  129. - whether every image is non-decorative and has a named semantic job;
  130. - absence of generic AI visual clichés used only as decoration;
  131. - coherence of paper, ink, grain, tape, shadow, or other material treatments;
  132. - sequence-level image cadence: no accidental return to a wall of text after an image-led opening;
  133. - semantic fit between each page and its recorded composition intent;
  134. - meaningful variation from recent decks without random novelty;
  135. - at least one deliberate moment of contrast or surprise that serves the argument.
  136. Diagram budget conformance is suggested, not blocking: `references/diagram-system.md` 的复杂度预算(节点数、字号档、宽度档)是默认参考,内容关系需要时允许偏离,偏离要在 manifest 里写一句理由。理解失败永远优先于预算合规。
  137. Read the three statuses separately: `structurally_valid` means the page did not break; `veg_review_required` remains true until review; `visually_approved` is never granted by the renderer. Do not accept a deck that merely passes overflow checks. Visual balance, awkward orphan lines, misleading emphasis, and a manifest that the render fails to express still require judgment. Revise the JSON and rerender until both automated and visual checks pass.
  138. After the last content, SVG, theme, or layout change, rerender the whole deck and inspect every PNG again. Write `qa-summary.md` using the review contract. A previous PASS becomes stale as soon as any source changes.
  139. ## 8. Deliver the release bundle
  140. Run the fail-closed delivery gate after final visual QA:
  141. ```bash
  142. python3 <skill-dir>/scripts/check_delivery.py carousel.json \
  143. --output-dir rendered \
  144. --qa-summary qa-summary.md
  145. ```
  146. Do not report a carousel `release` unless it exits `0` with `release_ready: true`. Otherwise report `blocked` or `candidate` and name the exact unmet gate.
  147. Return:
  148. - release state (`release`, `candidate`, or `blocked`);
  149. - the storyboard;
  150. - the editable JSON spec and local source assets;
  151. - the HTML preview as the deterministic rendering layer;
  152. - the rendered PNG directory as the final product;
  153. - `qa.json`, `qa-summary.md`, and the delivery-check result;
  154. - the design manifest and any reviewed recent-similarity warning;
  155. - optional post caption and hashtags only when requested.
  156. Keep publishing separate. Uploading or posting to Xiaohongshu is an external action and requires explicit authorization at action time.
README.md · README · 186 lines
  1. # 画图 Skill
  2. 画图 Skill(huatu-skill)是一个面向 Codex 等支持 `SKILL.md` 的 AI Agent 的视觉叙事 Skill:帮内容创作者把长文、研究笔记和草稿,做成可编辑、可验证的小红书图文轮播(1080×1350 PNG 信息卡片),或一个自包含的 ELI5 一页式 HTML 视觉解释器;本地渲染,交付前有 fail-closed 质检。
  3. English: an agent skill (`SKILL.md`) that turns articles and research into Xiaohongshu (RedNote) carousel PNGs or self-contained ELI5 visual explainers, with local Playwright rendering and fail-closed QA.
  4. > Constraints keep the page safe. Interpretation makes it meaningful.
  5. ![画图 Skill 示例](docs/preview/slide-01.png)
  6. ## 它解决什么
  7. 画图 Skill 先理解论点、关系与情绪,再选择两种产品路径之一:小红书轮播输出可编辑 JSON、HTML 预览、1080×1350 PNG 与 QA;ELI5 Visual Explainer 输出一个自包含响应式 HTML 与浏览器验证证据。**轮播的最终产品是 PNG;解释器的最终产品是 HTML。`--check-only` 对两者都只是预检**。
  8. - 将长文章、研究笔记或已有草稿改写成 6–10 页轮播叙事
  9. - 用可检查的 `designManifest` 记录视觉命题、隐喻、节奏与构图意图
  10. - 支持 editorial、mixed、image-led、photo-diary、object-study、poetic-poster 六种视觉模式
  11. - 知识型内容默认走 image-led / mixed 图文交替路径,并可复用 6 种 diagram SVG 模板表达流程、步骤、层级、时间、象限与对比关系
  12. - 支持本地 PNG、JPEG、WebP、SVG,并赋予 hero、evidence、atmosphere 等语义角色
  13. - 检查溢出、低对比度、重复构图、弱图像节奏与近期设计相似度
  14. - 生成本地 HTML 预览和逐页 PNG,并用 fail-closed delivery checker 阻止缺图、陈旧渲染或 BLOCKED QA 被当成成品
  15. - 生成 3–7 幕、低文案、大视觉的 ELI5 HTML 解释器,明确来源与简化边界
  16. - 在真实 Chromium 的 390×844 与 1365×768 视口检查外部请求、横向溢出、小字和浏览器错误
  17. - 把最终审美判断、手机验收与发布权留给人
  18. ## 适用场景与区别
  19. | | 说明 |
  20. |---|---|
  21. | 适合 | 知识型、研究型、机制型的小红书图文;把一篇文章拆成 6–10 页的叙事;需要一页式、手机和桌面都能看的解释页 |
  22. | 不适合 | 自动发帖或排期(本项目不做);纯照片修图;需要 JavaScript 交互的网页 |
  23. 和其他做法的区别:
  24. - 对比「直接让 AI 生图」:文字由 HTML/CSS 排版,产出可编辑的 JSON 源文件,再确定性渲染,所以中文排版可以修改、可以复现。
  25. - 对比「手工套模板」:先写叙事和设计意图(`designManifest`),再按内容选择构图。
  26. - 对比「普通 HTML 转图片工具」:多了溢出 / 对比度检查,以及 fail-closed 的交付门槛。
  27. ## 设计原则
  28. 1. 先写每页的叙事任务,再开始排版。
  29. 2. 图像必须承担意义,而不是给文字加一层装饰。
  30. 3. 规则是护栏,不是模板;内容变化应该带来构图变化。
  31. 4. 自动检查只能证明页面没有坏,不能证明它好看。
  32. 5. `--check-only` 不是交付;最后一次修改后必须重新渲染并逐页检查。
  33. 6. 不复制其他创作者的文字、水印、头像或标志性身份元素。
  34. ## 快速开始
  35. 1. 克隆到 skills 目录并安装渲染依赖:
  36. ```bash
  37. git clone https://github.com/ythx-101/huatu-skill.git ~/.codex/skills/huatu-skill
  38. cd ~/.codex/skills/huatu-skill
  39. python3 -m pip install -r requirements.txt
  40. python3 -m playwright install chromium
  41. ```
  42. 已经装了 Chrome/Chromium 的,可以用 `--browser <路径>` 或环境变量 `CHROME_BIN` 指定浏览器。其他支持 `SKILL.md` 的 Agent(例如 Claude Code 的 `~/.claude/skills/`)按各自的 skills 目录放置即可。
  43. 2. 在 Agent 里说一句话。轮播:
  44. ```text
  45. 使用 $huatu-skill,把这篇文章做成 8 页小红书图文。
  46. 先提出三种真正不同的视觉方向,选择后再渲染;加入有语义作用的图像,最后检查每一页和整组节奏。
  47. ```
  48. ELI5 解释器:
  49. ```text
  50. 使用 $huatu-skill,用 ELI5 HTML 模式把这篇文章做成 5 幕的一页式视觉解释器。
  51. ```
  52. 3. Skill 会先选择产品模式。轮播路径完成:内容契约 → 分页叙事 → JSON spec → 本地渲染 → 逐页视觉 QA → fail-closed 交付检查。解释器路径完成:3–7 幕叙事 → 单幕单概念 → 自包含 HTML → 双视口 Chromium QA → 人工检查。
  53. ## ELI5 Visual Explainer
  54. 从安全起始文件开始:
  55. ```bash
  56. cp assets/eli5-explainer-starter.html explainer.html
  57. python3 scripts/check_explainer.py explainer.html --check-only
  58. python3 scripts/check_explainer.py explainer.html --output-dir explainer-qa
  59. ```
  60. 真实浏览器命令生成 `qa.json`、`preview-mobile.png` 和 `preview-desktop.png`。HTML 只允许内联 CSS/SVG;不允许 JavaScript、远程字体/CDN、分析代码、网络资源、凭据、私有标识或二维码。参见 `references/eli5-html-mode.md`,以及可复现示例 `examples/eli5-moshi-hook.html` 和 `examples/eli5-moshi-browser-preview.html`。
  61. ## 轮播本地渲染
  62. 先检查结构:
  63. ```bash
  64. python3 scripts/render_carousel.py examples/mixed-carousel.json \
  65. --output-dir rendered --check-only
  66. ```
  67. 再生成 HTML 和 PNG:
  68. ```bash
  69. python3 scripts/render_carousel.py examples/mixed-carousel.json \
  70. --output-dir rendered --no-history
  71. ```
  72. 输出包括:
  73. - `carousel.html`:可编辑的本地预览
  74. - `slide-01.png` 等:1080×1350 图片
  75. - `qa.json`:尺寸、溢出与诊断结果
  76. - `design-manifest.json`:可检查的设计意图
  77. 渲染 PNG 需要 Playwright 与本地 Chromium(见「快速开始」);仅做 JSON 结构检查不需要浏览器,但结构检查不能替代成品验收。
  78. 最终逐页检查后,写入 `qa-summary.md`,再运行:
  79. ```bash
  80. python3 scripts/check_delivery.py examples/mixed-carousel.json \
  81. --output-dir rendered \
  82. --qa-summary qa-summary.md
  83. ```
  84. 只有输出 `release_ready: true` 才满足机械交付门槛;它仍不授权自动发布。
  85. ## 目录
  86. ```text
  87. SKILL.md Skill 工作流
  88. agents/openai.yaml Codex 展示信息
  89. assets/ 轮播与 ELI5 HTML 起始模板
  90. references/ 叙事、构图、图像、交付与 QA 规范
  91. scripts/render_carousel.py 轮播校验与渲染器
  92. scripts/check_delivery.py 轮播成品包 fail-closed 交付检查
  93. scripts/check_explainer.py ELI5 HTML 静态与双视口浏览器检查
  94. scripts/compare_directions.py 候选方向差异检查
  95. tests/ 单元测试
  96. examples/ 可复现示例
  97. ```
  98. ## 测试
  99. ```bash
  100. python3 -m unittest discover -s tests -v
  101. ```
  102. CI 中不运行浏览器测试(未安装 Playwright 时这些测试会自动跳过)。
  103. ## FAQ
  104. **能直接发布到小红书吗?**
  105. 不能。Skill 不登录、不上传、不自动发布,发布需要人工操作。
  106. **需要 API Key 或联网吗?**
  107. Skill 自身的脚本在本地渲染,不需要额外的 API Key;除首次安装 Playwright 与 Chromium 需要下载外,渲染和检查不依赖网络。图片只接受本地文件,远程 URL 会被拒绝。Agent 本身所用模型如何计费,不在本项目范围内。
  108. **没装 Playwright / Chromium 能用吗?**
  109. 只能跑 `--check-only` 结构预检,这不算交付;要渲染 PNG 或做浏览器 QA,必须装上。
  110. **支持哪些尺寸?**
  111. 默认 1080×1350(4:5)。单张 3:4 封面可以设 `canvas: {"width":1080,"height":1440}`;宽高允许范围 720–2160。
  112. **能改配色吗?**
  113. 默认是 Kami 主题,可以在 spec 里用 `theme` 覆盖(六位 hex 色值),见 `references/kami-design.md` 与 `references/spec-format.md`。
  114. **支持哪些视觉模式?**
  115. editorial、mixed、image-led、photo-diary、object-study、poetic-poster。
  116. **能在 Claude Code 或其他 Agent 里用吗?**
  117. 它是标准的 `SKILL.md` 结构,可以放进对应 Agent 的 skills 目录;本文示例按 Codex 编写。
  118. **ELI5 HTML 里能放 JavaScript 或外链字体吗?**
  119. 不能,只允许内联 CSS/SVG。
  120. **`release_ready: true` 是不是就代表图好看?**
  121. 不是。它只证明机械门槛通过,审美判断和发布权在人手里。
  122. **可以用别人的图吗?**
  123. 只使用原创、已授权或公有领域的素材。
  124. ## 隐私与版权
  125. - 图片只接受本地文件路径,远程 URL 会被拒绝。
  126. - Skill 不登录、上传或自动发布小红书;发布是独立的人工动作。
  127. - 请只使用原创、已授权或公有领域素材,并自行核对事实与来源。
  128. ## License
  129. MIT © 2026 QingYue

Ratings and comments

Sign in to rate this skill.

Sign in

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