小红书卡片视觉生产 · 创作技能

这个技能解决什么问题

把已批准的小红书草稿与真实产品素材做成可审查卡片：先完成 DAI 与 image plan 并锁定文案，再按确定性、可分离图层或完整效果三种模式制作，机械验证 PNG 与清单并运行合规检查。精确文字与真实 UI 不交给模型猜测，AI 生图需单独授权。

适用任务

为已批准的小红书草稿生产成套卡片图：锁定文案与证据、规划图片、选择制作模式、先做三张竖切片审查，再扩展并运行验证脚本与合规检查。

需要的输入

必需：已批准草稿与逐字卡片文案、账号角色、读者与一句话用户价值；经核实的产品事实与产品自有截图或品牌素材；输出目录（位于调用方 memory/xhs-replicate/ 运行内）；目标画布、卡片数量、版权与隐私边界、是否批准付费生图；使用外部视觉参考时需一份已校验的 visual-reference-selection.json。

交付内容

image-plan.md、visual-manifest.json、render/ 下经审查的 PNG、visual-review.md（含来源、所选模式、生成资产、替换决策、尺寸、隐私、版权、事实与手机可读性结论）；验证脚本与 make lint-post 的检查结果。

任务示例

用 xhs-render-cards 为 posts/xhs-ai-notes/ 的已批准草稿做 6 张卡片：先出封面、方法清单、产品截图三张竖切片给我看，用确定性模式，不启用 AI 生图。

输出示例

示意（非真实运行结果）：image-plan.md 列出每张卡片的文案与证据来源；render/ 下生成三张 PNG 并在对话中展示；visual-manifest.json 记录尺寸与模式；visual-review.md 记录审查结论；验证脚本输出通过或具体失败项。

来源与使用

作者 tsingyuai；许可 Apache-2.0。已核对来源与包结构，未据此宣称完成运行实测。

查看原始来源

机器可读详情 · 读取 SKILL.md · 下载完整技能包

让 Agent 使用这个技能 ↗

技能说明全文

---
name: xhs-render-cards
description: "把已批准的小红书草稿、单一分析参考和真实产品素材变成可审查卡片：先完成 DAI 与 image plan，再按确定性、完整效果或可分离图层模式制作，机械验证 PNG 与清单并运行合规检查。精确文字和真实 UI 不交给模型猜测；AI 生图需要单独配置和授权。"
---

# Xiaohongshu card production

Read [visual-production-contract.md](references/visual-production-contract.md), [rendering.md](references/rendering.md), and [visual-review-rubric.md](references/visual-review-rubric.md). Keep DAI and compliance rules from [deai.md](references/deai.md) and [quality-gates.md](references/quality-gates.md).

## Required inputs

Require before rendering:

- approved draft, exact card copy, account role, reader, and one-sentence user value;
- verified Product facts and Product-owned screenshots or brand assets;
- one validated `visual-reference-selection.json` when external visual research is used;
- output directory in the calling `memory/xhs-replicate/` run;
- target canvas, card count, rights/privacy boundaries, and whether paid image generation is approved.

The selected public reference is analysis-only. Learn hierarchy, proof-zone proportion, density, and reading rhythm; never copy its wording, logo, proprietary UI, exact composition, distinctive decoration, or full sequence. Rejected candidates contribute no visual rules.

## 1. Lock copy and evidence

Run DAI on title, caption, card copy, CTA, and tags. Every Product claim needs a current evidence source. Delete unsupported claims instead of weakening them with defensive filler.

Write `image-plan.md` using [image-plan.md](references/image-plan.md). Lock every visible string before rendering. Do not add cards merely to match the reference.

## 2. Capture real Product evidence

Use [screenshot-assets](../screenshot-assets/SKILL.md) for real Product UI, website, code, or data. Keep the original screenshot as evidence. Do not ask an image model to recreate Product UI, logos, citations, statistics, or source text and present it as real.

## 3. Choose one production mode

- `deterministic`: HTML/CSS, Canvas, SVG, or Pillow owns layout and exact text. Preferred for Chinese copy, dense diagrams, real screenshots, and repeatable series.
- `separable-layer`: the image model creates only a background, texture, or illustration without text/UI; deterministic rendering places approved copy and Product evidence.
- `complete-effect`: the image model proposes a whole visual direction. Generated text, UI, logos, evidence, and sensitive claims must be replaced or the candidate rejected.

Image generation is optional and separately authorized. If neither `OPENAI_API_KEY` nor `GEMINI_API_KEY` is configured, show the user the exact steps in [`CONFIGURATION.md`](../../CONFIGURATION.md) and offer deterministic rendering or a no-image handoff. Never silently retry or request a key in conversation.

## 4. Test a three-card vertical slice

Before expanding a large pack, render:

1. hook/cover;
2. independently useful method/checklist/comparison;
3. real Product evidence when the content makes a Product claim.

Show the actual images in the conversation. A filesystem path alone is not a visual review. Check phone-size readability, hierarchy, Product truthfulness, privacy, and copying boundaries.

## 5. Validate and review

Create `visual-manifest.json`, then run:

```powershell
python executors/xhs-render-cards/scripts/validate_social_card_pack.py --package <visual-package>
```

Run the existing DAI and compliance checks with `make lint-post POST=<post-directory>`. Inspect every final image at full size and as a contact sheet. Write `visual-review.md` with sources, selected mode, generated assets, replacement decisions, dimensions, privacy, copyright, factual and mobile-readability results.

Only reviewed files under `render/` may enter the human publishing package. Rendering never authorizes upload or publication.

## Stop rules

Stop on missing Product evidence, ambiguous rights, unreviewed reference selection, exposed private data, fake UI, malformed text, provider failure, or repeated candidate failure. One targeted retry may replace an explicitly authorized work-in-progress candidate; do not loop or create duplicate publishable assets.

技能是任务材料，不会自动授予账户权限、付费授权或其他系统权限。请让 Agent 按说明核对依赖，再在授权范围内使用。