SHILIU · 创作资源

小红书只读选题与视觉参考采集 · 创作技能

通过本机 browser-first xiaohongshu-mcp 只读搜索小红书笔记，批量下载候选首图并按推广版式质量评分，补全用户选定笔记详情，把脱敏证据写入调用方Memory，用于选题与视觉参考研究。

拾流编辑部 · 更新于 2026-11-02

这个技能解决什么问题

适用任务

围绕给定主题只读采集小红书笔记与首图，筛选视觉参考样本，输出脱敏的选题与视觉证据供后续创作使用。

需要的输入

先解压完整包，按 SHILIU-INSTALL.md 读取 upstream/collectors/xiaohongshu-mcp/SKILL.md。主题关键词；本机 xiaohongshu-mcp 服务（本地HTTP）；已登录的小红书账号（登录仅授权只读采集）；批次数量（建议25）；输出目录 memory/xhs-replicate/<run>/；CONFIGURATION.md 中要求的配置项。

交付内容

xiaohongshu-search.json 搜索证据、候选首图与contact sheet、visual-candidates.json、visual-reference-selection.json（经校验脚本验证）、笔记详情摘要、交接说明（查询、批次、时间、访问限制、候选、公开URL、缺失证据、下一步建议）。

任务示例

请用小红书只读采集帮我研究“通勤穿搭”选题，先采25条笔记并下载首图，按推广版式质量评分后给我3-8个候选，附公开链接，我选定后再补详情。

输出示例

示意结构（非真实结果）：一、采集参数（查询词、批次25、时间、访问限制）；二、候选列表（每条含首图、标题、版式评分、风险、公开URL）；三、用户选定后的详情摘要；四、visual-reference-selection.json 校验通过记录；五、缺失证据与下一步建议。

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

技能说明全文

---
name: xiaohongshu-mcp
description: 使用本机 browser-first xiaohongshu-mcp 只读搜索小红书、下载候选首图、补全用户选择的笔记详情，并把脱敏证据写入调用方 Memory。用于 xhs-replicate 的选题和视觉参考研究，取代小红书 MediaCrawler 路径。
---

# Xiaohongshu browser-first collection

Read [runtime.md](references/runtime.md) before startup and [cover-screening.md](references/cover-screening.md) before visual selection.

## First-run conversation

Before collection, tell the user:

- the recommended first-run batch is 25 notes;
- the count is adjustable, but 25 is recommended;
- collection is read-only and saves sanitized research evidence and requested images locally;
- login does not authorize likes, saves, comments, follows, uploads, or publication.

If required settings are missing, stop and invoke `onboard-growth-lab`. Give the user the exact configuration file and fields from [`CONFIGURATION.md`](../../CONFIGURATION.md); never ask them to paste a key, cookie, or signed URL into the conversation.

## Runtime

```powershell
powershell -ExecutionPolicy Bypass -File collectors/xiaohongshu-mcp/scripts/start_xiaohongshu_service.ps1
python collectors/xiaohongshu-mcp/scripts/collect_xiaohongshu.py "<topic>" `
  --limit 25 --cover-pool 25 `
  --out "memory/xhs-replicate/<run>/xiaohongshu-search.json"
```

The service must be local HTTP only. If it is not logged in, explain the boundary, ask before opening the visible login window, run `login_xiaohongshu.ps1`, verify once, and resume. Stop on timeout, risk-control, login loss, or repeated empty responses; do not loop around platform controls.

## Visual selection

1. Persist the 20-30 item search response immediately as one batch. Do not wait for page-wide stability after the response is complete.
2. Download all covers from the first batch and inspect every contact sheet.
3. Score promotional layout quality before engagement. Fetch full details only for 3-8 passing candidates.
4. Show every passing candidate with its actual representative image, title, score, and risk. If inline image rendering is unavailable or cannot be confirmed, include the clean public note URL in the same response.
5. If the user rejects all candidates, record the reasons and run a new product/workflow-oriented query. Do not force the best item from a weak batch.
6. Select exactly one external visual learning sample. Write and validate `visual-reference-selection.json`:

```powershell
python collectors/xiaohongshu-mcp/scripts/validate_visual_reference_selection.py `
  --selection <run>/visual-reference-selection.json `
  --candidates <run>/visual-candidates.json
```

Use the selected reference for analysis only. Do not copy its wording, logo, proprietary UI, exact composition, or visual identity.

## Data boundary

Write search evidence, candidate images, selection, and detail summaries only under the calling Model's ignored `memory/xhs-replicate/` run directory. Persist clean note IDs and public URLs, never xsec tokens, cookies, signed media URLs, avatars, or raw response fields containing credentials.

## Handoff

Return the query, batch size, collection time, access limits, candidates, visible engagement, selection status, clean public URLs, missing evidence, and one recommended next action. The user must be able to see the candidate image or public source before choosing.

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。