猎头人才库通用启动 Skill

This is the general member version. Its job is to help a recruiter start smoothly, avoid setup/browser/Feishu pitfalls, and then ask what work the recruiter wants done.

It must not force Tommy/zzclaw's private sheet columns or private outreach rules onto other recruiters. Always adapt to the member's own table headers and industry.

First Response

When the user says "按照这个 skill 执行", "帮我录入简历", "初始化", "安装工具", or gives this skill to a newly installed WorkBuddy/QClaw, start with:

我先帮你检查环境是否能直接读取飞书表格。检查通过后再开始录入简历。

我可以帮你把候选人简历录入到飞书表格。你先把要录入的飞书表格链接发给我。

常见录入方式有三种：
1. 你在飞书表格某一列放入候选人的脉脉/领英个人主页链接。我会在你登录账号后，逐个打开链接，提取候选人信息，并写回表格。
2. 你在飞书表格某一列插入简历附件。我会逐个读取附件内容，并按表头录入。
3. 你告诉我电脑上简历所在的文件夹路径。我会打开这些本地简历，提取信息后录入表格。

录入时，我可以顺手帮你做候选人岗位方向分类，默认按“一级类目-二级类目-三级类目”的方式填写。如果你有自己的分类习惯，提前告诉我，我会按你的标准来。

录入完成后，如果你需要，我还可以继续帮你在脉脉/领英上加好友、发私信，或者准备邮件。

以上是固定常用场景。除此之外，如果你还有其他和候选人整理、人才库维护、沟通跟进相关的需求，也可以直接告诉我。

Do not ask beginners to choose technical tools. The agent chooses the path.

Mandatory Environment Doctor

For every new member machine, and before any Feishu table reading/writing, run:

python scripts/env_doctor.py --url "<feishu url>"

If no Feishu URL is available yet, ask for the Feishu table link first. Do not process resumes, open Maimai/LinkedIn, use bot API, or inspect Feishu in a browser before the doctor returns ready.

The only allowed statuses before work starts are:

• ready: continue with lark-cli --as user.
• need_one_terminal_command: show the member only the one command from member_action; do not add extra explanation or alternative commands.
• need_auth_login: show only the one auth command from member_action and ask the member to finish Feishu authorization.
• need_admin_permission: stop and explain this is a table/app permission issue, not a member computer issue.
• need_install_tools: run setup/install only after the member agrees; if install fails, escalate to the assistant/admin.
• blocked_unknown: stop and ask the member to send the diagnosis to the assistant/admin.

When a status is not ready, do not continue experimenting. Do not use bot API, browser DOM/XHR/canvas extraction, CSV export, internal Feishu bridge objects, or link-sharing permission changes as fallback paths.

Required References

Read these before acting:

• references/member-workflow.md for the startup and source routing flow.
• references/tooling.md for lark-cli, local files, browser/CDP, Windows ACP, and login persistence.
• references/classification.md whenever the table has role/category/direction fields or the user asks to tag candidates.

Core Principles

• First run setup before work: check tools, Feishu auth, browser path, and local file support.
• Environment diagnosis is a gate, not a suggestion. If env_doctor.py is not ready, handle that status first and do not start candidate work.
• If Feishu auth fails, separate authorization, local credential storage, and host execution permission. Do not diagnose Keychain/sandbox errors as Chrome, browser login, or sheet-row problems. On macOS, read references/tooling.md before asking the user to reauthorize repeatedly.
• If lark-cli user auth/token storage is blocked by sandbox, Keychain, or ~/Library/Application Support, stop there and fix user credential storage first. Do not fall back to bot identity, do not ask the member to open document link sharing, and do not ask them to add an unknown Feishu app as collaborator just because bot API returns 91403 Forbidden.
• Always read the member's table headers before writing.
• Map fields by meaning, not by fixed column letters.
• If a table lacks a suitable field, ask where to put the information or leave it out; do not write into unrelated columns.
• Process web profile links one by one: open one link, read/expand, write one row, verify, then continue.
• Use browser automation only for web-only profile pages such as Maimai/LinkedIn. Do not use browser automation to read local Excel or Feishu data when API/local parsing is available.
• Prefer fixed logged-in Chrome/Edge + CDP port/profile for Maimai/LinkedIn. Avoid temporary xbrowser profiles that lose login.
• Reuse one recruiting tab/page. After the user logs in once, open each new Maimai/LinkedIn candidate by navigating/overwriting the existing page, not by launching a new browser or opening many tabs. Updated browser_cdp.js goto must reuse the same working tab by default; only use --new-tab for temporary diagnosis.
• Do not add friends, send messages, or send emails unless the user explicitly asks.
• Resume attachments/local resumes default to extraction and table entry only, not outreach.
• Changed cells should be highlighted when the table supports it, but only actual changed cells.
• Preserve existing real phone/email/WeChat/links unless new evidence proves they are wrong or the user asks to overwrite.

Structured Tagging

When the table has role/category/technical direction/business direction fields, tag candidates in a structured hierarchy:

一级-二级-三级

If only first/second level can be seen, stop there. Do not invent a third level.

For software/AI/technical candidates, use the detailed software/AI taxonomy in references/classification.md.

Example pattern:

算法-大模型算法-LLM-长任务规划与执行

If the member works in healthcare, manufacturing, finance, consumer, real estate, HR, or another industry, imitate the same 一级-二级-三级 structure with that industry's own role logic. Follow the member's own category list when they provide one.

For other industries, use common-sense industry categories in the same structure. Examples:

• 制造业-质量管理-供应商质量
• 制造业-生产管理-精益生产
• 金融-投研-量化研究
• 金融-风控-信用风险
• 消费-品牌市场-用户增长
• 医药-临床研究-CRA
• 地产-工程管理-机电工程

The purpose is not to be "scientifically perfect"; the purpose is to make the recruiter's talent table searchable, filterable, and reusable.

Public Search Trigger

Good-background candidates should trigger public search.

For software/AI/technical/research candidates, keep the technical search path when relevant: GitHub, Google Scholar, OpenReview, Semantic Scholar, DBLP, AMiner, personal homepage, arXiv, conference pages, technical blogs, and company/model-release pages.

For other industries, do not force technical academic paths. Search by candidate name + current/recent company name, then add title, school, city, industry keyword, project/case keyword, award, patent, article, interview, conference, or public event keyword as needed. The goal is to find more complete reliable public information and useful contact clues.

Output After Setup

After environment checks are complete, report:

环境状态：ready / local-only / blocked-by-host
已具备：...
还缺少/阻塞：...

现在你可以让我做：
1. 从飞书表格某一行开始录入；
2. 处理表格里的脉脉/领英链接；
3. 处理表格里的简历附件；
4. 处理本地 Excel / Word / PDF 简历；
5. 配置邮件或准备触达文案。

你把表格链接、行号、文件路径或候选人链接发给我即可。

If blocked, explain the category clearly and continue with any local work that is still possible.