dabi-chromeagent-relay 1.0.0

package/skills/dabi-relay-xiaohongshu/SKILL.md

@@ -0,0 +1,280 @@
+---
+name: dabi-relay-xiaohongshu
+description: Publish Xiaohongshu (小红书 / RED) notes and search/list Xiaohongshu posts through the dabi-relay plugin tools backed by the local ChromeAgent Relay service. Use when the user wants to 发小红书、发布图文、保存草稿、搜索小红书帖子、获取帖子列表、找对标内容、导出搜索结果，或检查小红书连接是否正常。 Prefer dabi-relay Xiaohongshu tools over generic browser automation when this plugin is available.
+---
+
+# Dabi Relay Xiaohongshu
+
+Use dabi-relay as the default path for Xiaohongshu work.
+
+When this skill applies, prefer the dedicated Xiaohongshu tools instead of generic browser automation.
+
+## Main goal
+
+Make Xiaohongshu actions feel simple and consumer-friendly.
+
+The user should feel like they are asking for:
+
+- 发小红书
+- 存草稿
+- 搜帖子
+- 导出结果
+- 看看连接有没有问题
+
+Do the technical selection internally.
+
+## Default tool preference
+
+Prefer these tools first:
+
+- `xhs_publish_note`
+- `xhs_save_note_draft`
+- `xhs_search_notes`
+- `export_to_csv`
+
+Use these only when needed for diagnosis or lower-level handling:
+
+- `chromeagent_relay_health`
+- `chromeagent_relay_list_tools`
+- `chromeagent_relay_execute_tool`
+- `chromeagent_relay_wait_task`
+- `xhs_list_runtimes`
+- `xhs_list_tabs`
+
+Do not default to the browser tool if dabi-relay can complete the task.
+
+## User-facing language
+
+Use plain language.
+
+Prefer saying:
+
+- 小红书连接
+- 浏览器连接
+- 插件连接
+- 连接正常
+- 还没连上
+- 我先帮你检查一下
+- 我继续帮你发
+
+Avoid these terms unless the user is explicitly debugging:
+
+- runtime
+- runtimeId
+- relay task id
+- tool schema
+- action task
+- clientId
+
+If you must use a lower-level diagnostic tool, do it silently and summarize in human language.
+
+## Intent routing
+
+### 1) Publish now
+
+Use `xhs_publish_note` when the user clearly wants immediate posting.
+
+Typical phrases:
+
+- 帮我发小红书
+- 直接发布
+- 现在发出去
+- 帮我把这篇发掉
+
+Required inputs:
+
+- `title`
+- `content`
+- `images` as absolute local file paths
+
+Optional inputs:
+
+- `topics`
+
+Rules:
+
+- Keep title within 20 characters.
+- Keep topics to 10 or fewer.
+- Preserve the user’s wording unless asked to rewrite.
+- If the user did not clearly ask to publish, prefer draft over publish.
+- Do not ask the user to choose any internal session or identifier.
+- If required inputs are missing, ask one short follow-up message listing only what is missing.
+
+### 2) Save draft
+
+Use `xhs_save_note_draft` when the user wants preparation without publishing.
+
+Typical phrases:
+
+- 先存草稿
+- 先别发
+- 先保存一下
+- 帮我做成草稿
+
+Use the same validation rules as publish.
+
+### 3) Search or get a list of posts
+
+Use `xhs_search_notes` when the user wants discovery, examples, competitors, or note lists.
+
+Typical phrases:
+
+- 搜一下小红书关于 xx 的帖子
+- 获取帖子列表
+- 找一些对标内容
+- 抓一些样本
+- 看看别人怎么写
+- 帮我看看这些帖子的详情
+- 搜几篇帖子，把详情也抓下来
+- 抓 10 条笔记的完整内容
+
+When the user mentions 详情、完整内容、正文、评论、互动数据, set `detailTargetCount` to match the requested count (or default to `targetCount` if not specified).
+
+Useful inputs:
+
+- `keyword`
+- `targetCount`
+- `filters`
+- `detailTargetCount` — when the user wants detailed note info (title, content, comments, interactions), not just the search list
+- `randomizeOrder` — randomize the order of detail collection (default true)
+- optional legacy shorthands: `sort`, `noteType`
+
+Prefer `filters` over shorthand parameters.
+
+When `detailTargetCount` is provided, the crawler will first collect the search list, then open each note to capture full details (title, body, comments, interaction counts, etc.).
+
+Supported filter groups include:
+
+- 排序依据: 综合, 最新, 最多点赞, 最多评论, 最多收藏
+- 笔记类型: 不限, 视频, 图文
+- 发布时间: 不限, 一天内, 一周内, 半年内
+- 搜索范围: 不限, 已看过, 未看过, 已关注
+- 位置距离: 不限, 同城, 附近
+
+If the user asks for “获取列表”, return a concise summary first. Offer CSV export when it would help.
+
+### 4) Export results
+
+Use `export_to_csv` after search when the user wants:
+
+- 导出结果
+- 保存成 CSV
+- 发给同事分析
+- 留档
+
+Prefer a clear absolute output path.
+
+If the user did not give one, suggest a sensible path and ask once.
+
+Example:
+
+- `/Users/rui/Desktop/xiaohongshu-results.csv`
+
+### 5) Check whether the setup is working
+
+When the user asks why publishing/searching failed, first treat it as a connection problem, not a technical debugging session.
+
+Typical phrases:
+
+- 为什么发不了
+- 怎么没反应
+- 帮我检查一下
+- 看看连接是不是坏了
+
+Diagnosis order:
+
+1. `chromeagent_relay_health`
+2. `xhs_list_tabs`
+3. `xhs_list_runtimes` only if needed internally
+4. `chromeagent_relay_list_tools` only if tool exposure looks wrong
+
+Translate the result into plain language.
+
+Good examples:
+
+- 小红书连接正常，我继续帮你处理。
+- 浏览器扩展还没连上，所以现在还不能发。
+- 本地连接没起来，我先帮你检查一下。
+
+Bad examples:
+
+- runtime unavailable
+- no online runtime connected
+- action task failed
+
+## Ask only for what matters
+
+Ask follow-up questions only when essential information is missing.
+
+Typical missing items:
+
+- title
+- content
+- image absolute paths
+- publish vs draft intent
+- keyword for search
+- CSV output path
+
+Keep the follow-up short.
+
+Good example:
+
+- 把标题、正文和图片绝对路径发我，我直接帮你处理。
+
+Bad example:
+
+- 请确认 runtime、relay、tab、clientId、schema 是否正确。
+
+## Failure handling
+
+When something goes wrong, summarize by user-understandable category.
+
+Map technical issues into plain language:
+
+- relay not started → 本地连接没启动
+- extension not connected → 浏览器扩展没连上
+- no usable browser session → 浏览器这边还没准备好
+- invalid image path → 图片路径不对，或文件找不到
+- async action failed → 这次操作没成功，我可以重试或先帮你存草稿
+
+Do not dump raw JSON unless the user explicitly asked for debugging details.
+
+## Response style
+
+Be short, calm, and action-first.
+
+Preferred progress updates:
+
+- 我先检查一下连接
+- 我开始帮你发
+- 我先帮你存成草稿
+- 我先搜一批结果
+- 已经导出成 CSV
+
+Preferred result style:
+
+- 已发布
+- 草稿已保存，还没发布
+- 已搜到 30 条图文笔记
+- 连接正常
+- 现在还发不了，因为浏览器扩展没连上
+
+Avoid long technical explanations by default.
+
+## Practical defaults
+
+- If publish intent is ambiguous, choose draft.
+- If search scope is vague, use a moderate `targetCount` first.
+- If export is likely useful, offer it after summarizing search results.
+- If a lower-level tool returns an async action, wait properly instead of repeatedly retrying the original request.
+- Prefer one clean answer over multiple fragmented status messages.
+
+## Success standard
+
+The skill is working well when the user experience feels like this:
+
+- they ask in plain Chinese
+- the agent responds in plain Chinese
+- posting/search/export works without exposing internal terms
+- failures are explained in everyday language
+- the user never has to think about runtimes or bridge internals unless they explicitly want debugging