招聘 Skills · Recruiting Skills

中文 | English

给 Codex、Claude、Cursor 以及其他 Agent Skills 宿主用的研究者招聘技能包：从一篇 arXiv 论文补全全部作者的公开职业资料，或在本地索引里按毕业年份和论文门槛筛选人才队列，再只给入围者补证据。

这是招聘相关 Agent Skills 的长期集合。两个 Skill 可以单独安装，也可以按稳定候选人 ID 协作。仓库自带索引引擎和合成 fixture，不附带真实候选人数据库。

论文请求走第一个 Skill。形如“寻找 2027/2028 届、满足论文门槛的人”走第二个 Skill。筛选完成后，research-talent-scout 可以把入围候选人一次性交给 researcher-recruiting-skill 补全公开职业资料。

两个 Skill 如何协作

单篇论文
  └─ researcher-recruiting-skill
       ├─ 获取 arXiv 官方作者顺序
       ├─ 并行核验公开职业资料
       └─ 生成单文件 HTML 报告

批量人才条件
  └─ research-talent-scout
       ├─ 查询本地 SQLite 索引
       ├─ 按会议、一作类别、学位和毕业年份筛选
       ├─ 仅将入围候选人交给 researcher-recruiting-skill
       └─ 合并为一份证据化结果

research-talent-scout 负责索引内发现、确定性筛选、毕业与论文核验，以及最终合并结果。researcher-recruiting-skill 只对收到的稳定候选人 ID 做有界的公开资料补全。

交接使用唯一 batch_id、稳定 candidate_id 和明确的 deadline_at。资料补全不会扩展到论文的其他合作者，也不能改写原有论文统计、毕业年份状态或更强的既有证据。

仓库约定

• 克隆目录是 recruiting-skills。根目录 Skill 的安装名仍是 researcher-recruiting-skill，以兼容现有触发方式。
• 其他招聘 Skill 使用独立的 /SKILL.md 子目录。
• 每个 Skill 可以独立安装、触发和验证。
• 跨 Skill 协作必须定义稳定输入、输出、责任边界和失败降级。
• 通用合规原则保持一致：只处理公开职业信息，不执行招聘排名、自动雇佣决策或自动外联。

新增 Skill 后，需要在顶部目录表、安装说明和统一验证脚本中登记。发现用的仓库简介和 GitHub topics 写在 REPO_META.md；这些设置不能靠 git 文件自动生效。2026-10-29 的对照（可安装 SKILL.md、演示报告、双语 README、topics、MIT、CI，以及当时尚未套用的 topics）见 docs/trending-notes-2026-10-29.md。给代理看的短索引是根目录 llms.txt。

功能

单篇论文作者调查

• 接受 arXiv URL、arXiv ID 或精确论文标题；
• 保留 arXiv 官方顺序的完整作者列表；
• best-effort 查找机构、个人主页、公开职业邮箱、GitHub 和 OpenReview；
• 每个非空字段都附可点击的公开证据；
• 使用 已核验、可能匹配、暂未找到、未执行网页检索 状态；
• 默认资料补全预算约 20 秒；
• 生成 outputs//report.html，该目录已加入 .gitignore；
• 授权后可发布至 GitHub Pages 的 gh-pages:/index.html。

批量人才筛选

• 按博士/硕士、毕业年份、院校、实验室、会议和作者类别筛选；
• 支持 main_first、co_first、findings、workshop、short、demo 分类计数；
• 正确区分“多于 5 篇”和“至少 5 篇”；
• 将毕业年份分为 confirmed、likely_needs_verification、unknown、stale 和 conflict；
• 不会仅凭姓名合并同名研究者；
• 返回全部索引内匹配者，并显示索引覆盖和新鲜度限制；
• 使用 SQLite 进行快速查询。可以用 OpenAlex 或 Semantic Scholar 的公开元数据 API 写入 JSONL 再导入；仓库不附带真实候选人数据库。

安装

git clone https://github.com/LofiSu/recruiting-skills.git
cd recruiting-skills

链接名必须与 SKILL.md 的 name 一致。把下面的 dest 换成你使用的宿主目录；只安装需要的宿主即可。Cursor 也会读取 Codex 与 Claude 的技能目录。

dest="${CODEX_HOME:-$HOME/.codex}/skills"
# dest="$HOME/.claude/skills"
# dest="$HOME/.cursor/skills"
# dest="$HOME/.agents/skills"
# dest="/path/to/another-project/.cursor/skills"

mkdir -p "$dest"
ln -s "$PWD" "$dest/researcher-recruiting-skill"
ln -s "$PWD/research-talent-scout" "$dest/research-talent-scout"

只做单篇论文作者调查时，创建第一个链接即可。批量筛选和自动资料补全需要两个 Skill 都能被宿主发现。安装后请开启新一轮 Agent 会话以刷新 Skill 列表。不要把本仓库链接进它自己的 .cursor/skills/，那会把目录嵌套进自身。

使用示例

调查一篇论文的全部作者

https://arxiv.org/abs/2606.19980
帮我找这篇论文所有作者的机构、主页、公开职业邮箱和证据链接。

显式调用：

Use $researcher-recruiting-skill on https://arxiv.org/abs/2606.19980.
Find every author's institution, homepage, public professional email,
GitHub/OpenReview profile, and evidence.

批量筛选研究人才

Use $research-talent-scout to find indexed 2027 and 2028 graduates.
For PhD candidates require more than 5 qualifying papers; for master's
candidates require more than 3. Count ACL, ICML, ICLR, NeurIPS, EMNLP,
CVPR, AAAI, ICCV, MLSys, OSDI, EuroSys, ASPLOS, ISCA, SC, and PPoPP.
Separate confirmed graduation years from candidates needing verification,
then enrich only the qualifying candidates with public professional evidence.

“全部候选人”表示当前索引覆盖范围内的全部匹配者，不代表从实时互联网穷举全球所有研究者。

本地索引快速开始

演示数据是合成记录 research-talent-scout/tests/fixtures/records.jsonl（Alice Example、Bob Example 等），只用于演示和测试。

cd research-talent-scout

python3 scripts/init_index.py --db data/candidates.sqlite
python3 scripts/import_records.py \
  --db data/candidates.sqlite \
  --input tests/fixtures/records.jsonl

python3 scripts/query_candidates.py \
  --db data/candidates.sqlite \
  --graduation-years 2027,2028 \
  --degree-threshold phd=5,masters=3 \
  --strictly-greater \
  --venues ACL,ICML,ICLR,NeurIPS,EMNLP,CVPR,AAAI,ICCV,MLSys,OSDI,EuroSys,ASPLOS,ISCA,SC,PPoPP \
  --categories main_first,co_first,findings,workshop,short,demo \
  --graduation-states confirmed,likely_needs_verification

data/*.sqlite 已被忽略，不会进入版本库。

从 OpenAlex / Semantic Scholar 导入

连接器在 research-talent-scout/scripts/connectors/，入口是 scripts/fetch_records.py。它们只读公开学术元数据，写成上面的 JSONL。不猜测邮箱，不下载全文，不把毕业年份标成 confirmed。--degree 是调用者给索引用的标签，不是 API 事实。identity_state: verified 只表示这一行对应一个来源作者 ID，不是雇佣核验。同名作者不会被合并。

默认请求间隔：OpenAlex 0.2 秒（低于无 key 的每秒 10 次，也低于免费 key 的每秒 30 次）；Semantic Scholar 1.1 秒（按 introductory key 的每秒 1 次）。收到 HTTP 429 或 503 时最多重试 3 次，并遵守 Retry-After。

cd research-talent-scout
export OPENALEX_API_KEY=...   # 不要提交到 git
export S2_API_KEY=...         # 可选

python3 scripts/fetch_records.py \
  --source openalex \
  --venue NeurIPS \
  --year 2024 \
  --degree phd \
  --output data/openalex.jsonl

python3 scripts/fetch_records.py \
  --source semanticscholar \
  --author-id 1234567 \
  --degree phd \
  --output data/s2.jsonl

python3 scripts/init_index.py --db data/candidates.sqlite
python3 scripts/import_records.py --db data/candidates.sqlite --input data/openalex.jsonl

也可以按作者姓名搜索（--author）。多个来源 ID 会各成一行。Semantic Scholar 的论文检索按相关度排序，不是某个会议的全量普查，默认最多 25 篇。一作且 venue 能对上支持列表时记为 main_first；venue 字符串写明 workshop、findings、short 或 demo 时分开记；其余作者记为 unclassified。不推断共同一作。

离线样例是合成作者 Ada Fixture 和 Ben Fixture，不是真人库：

python3 scripts/import_records.py \
  --db data/candidates.sqlite \
  --input examples/sample-records.jsonl

演示报告

静态样例：examples/demo-report/index.html。它用现有报告模板填入公开论文 arXiv:1706.03762《Attention Is All You Need》的官方作者顺序，以及论文 HTML 里写明的机构。联系路径全部是「暂未找到」，没有邮箱，也没有 mailto 链接。

在浏览器中打开该文件即可看版式。GitHub 会把 HTML 显示成源码。用下面的命令可以按模板重新生成，并检查已提交的文件是否一致：

python3 scripts/render_demo_report.py
python3 scripts/render_demo_report.py --check

这个样例留在 main 上，方便审阅。CI 不会把它发布到 gh-pages。

验证

从仓库根目录运行：

bash tests/verify-all.sh

它会验证论文作者 Skill 的静态契约、两个 Skill 的 frontmatter、Pages 首页生成器、演示报告，以及 research-talent-scout 的索引、导入、查询、连接器映射和报告测试。连接器测试使用仓库里的 HTTP fixture，不访问网络。GitHub Actions 在推送到 main 和指向 main 的 pull request 上运行同一命令。frontmatter 检查使用仓库内的标准库脚本，不依赖某台机器上的 skill-creator 路径。

GitHub Pages

单篇论文模式生成的报告不会写入 main。授权后，报告会发布至：

branch: gh-pages
file:   /index.html
URL:    https://.github.io///
home:   https://.github.io//

本仓库对应 https://lofisu.github.io/recruiting-skills/。首页由 scripts/build_pages_index.py 和 assets/pages-index-template.html 生成。如果 GitHub 写入、Pages 设置或部署验证不可用，Skill 只返回本地报告，不会虚构线上地址。

想先看报告长什么样，打开仓库里的 examples/demo-report/index.html。那是提交在 main 上的静态样例，不是 gh-pages 上的一篇已发布报告。

数据与合规边界

• 只收集公开职业信息；
• 不根据姓名或机构规则猜测邮箱；
• 不登录、不绕过访问控制、不解 CAPTCHA、不批量抓取受限平台；
• 搜索摘要只能用于发现页面，不能单独作为已核验证据；
• 同名身份存在歧义时保持 可能匹配 或独立记录；
• 输出用于人工复核，不进行候选人排名、自动雇佣决策或自动外联。

贡献方式见 CONTRIBUTING.md。社区行为准则见 CODE_OF_CONDUCT.md。安全问题按 SECURITY.md 私下报告，不要开公开 issue。许可证为 MIT。

---

English

Agent Skills for Codex, Claude, Cursor, and any other host that discovers SKILL.md: enrich every author of an arXiv paper with public professional evidence, or screen an indexed graduate cohort and enrich only the qualifiers.

This is the long-term home for recruiting skills. The two skills install independently and collaborate through stable candidate IDs. The repository ships the index engine and synthetic fixtures. It does not ship a live candidate database.

Use the first Skill for a paper-centric request. Use the second for a cohort request such as “find 2027/2028 graduates above these publication thresholds.” After filtering, the scout hands exactly one qualifying-candidate batch to the recruiting Skill for bounded public-profile enrichment.

How they collaborate

research-talent-scout owns indexed discovery, deterministic filtering, graduation and publication verification, and the single final result. researcher-recruiting-skill owns only bounded public-profile enrichment for the stable candidate IDs it receives.

The handoff uses a unique batch_id, stable candidate_id values, and an absolute deadline_at. Enrichment never expands to coauthors, re-filters the cohort, or overwrites stronger graduation or publication evidence.

Repository convention

• Clone the repository as recruiting-skills. The root Skill install name remains researcher-recruiting-skill so existing triggers keep working.
• Additional recruiting Skills live in independent /SKILL.md directories.
• Every Skill must remain independently installable, triggerable, and verifiable.
• Cross-Skill workflows must define stable inputs, outputs, ownership, and failure behavior.
• Shared compliance rules permit only public professional information and prohibit candidate ranking, automated employment decisions, and automated outreach.

Register every new Skill in the catalog table, installation instructions, and tests/verify-all.sh. Suggested GitHub topics and the About description live in REPO_META.md; GitHub does not apply those settings from a file. The 2026-10-29 map of installable SKILL.md files, the demo report, the bilingual README, topics, MIT, and CI — including topics that were still unset — is docs/trending-notes-2026-10-29.md. The short index for agents is llms.txt at the repository root.

Features

Paper author investigation

• accepts an arXiv URL, identifier, or exact paper title;
• preserves the complete official author order;
• finds best-effort affiliations, homepages, public work email, GitHub, and OpenReview profiles;
• requires a clickable public source for every non-empty field;
• uses explicit verified, possible-match, not-found, and no-search states;
• targets approximately 20 seconds for native web enrichment;
• writes outputs//report.html (gitignored) and can publish it to GitHub Pages when authorized.

Cohort talent scouting

• filters by degree, graduation year, institution, lab, venue, and authorship category;
• counts main first, evidenced co-first, Findings, workshop, short, and demo papers separately;
• distinguishes “more than N” from “at least N”;
• preserves confirmed, inferred, unknown, stale, and conflicting graduation evidence;
• keeps ambiguous same-name researchers separate;
• returns all matches inside declared index coverage and reports freshness limitations;
• uses a lightweight SQLite index for fast repeated queries;
• can fill that index from the OpenAlex and Semantic Scholar public metadata APIs.

The connectors write normalized JSONL. They do not ship a live candidate database, and they do not mark graduation years as confirmed.

Installation

git clone https://github.com/LofiSu/recruiting-skills.git
cd recruiting-skills

The symlink name must match the name in each SKILL.md. Point dest at the host you use. Cursor also loads Codex and Claude skill directories, so a Codex or Claude install is visible in Cursor as well.

dest="${CODEX_HOME:-$HOME/.codex}/skills"
# dest="$HOME/.claude/skills"
# dest="$HOME/.cursor/skills"
# dest="$HOME/.agents/skills"
# dest="/path/to/another-project/.cursor/skills"

mkdir -p "$dest"
ln -s "$PWD" "$dest/researcher-recruiting-skill"
ln -s "$PWD/research-talent-scout" "$dest/research-talent-scout"

Install only the first link for paper author investigation without cohort filtering. Restart the Agent session after installation so Skill discovery refreshes. Do not link this repository into its own .cursor/skills/ directory.

Usage

Paper mode:

Use $researcher-recruiting-skill on https://arxiv.org/abs/2606.19980.
Find every author's institution, homepage, public professional email,
GitHub/OpenReview profile, and evidence.

Cohort mode:

Use $research-talent-scout to find indexed 2027/2028 PhD and master's
candidates above configurable first-author venue thresholds. Keep confirmed
and pending graduation evidence separate, then enrich only the qualifiers with
public professional evidence.

“All candidates” means all matches within declared index coverage, not an exhaustive live-web census of every researcher.

Populate the index from OpenAlex or Semantic Scholar

The clients live in research-talent-scout/scripts/connectors/. scripts/fetch_records.py writes JSONL for import_records.py. It reads public scholarly metadata only: no email guessing, no full-text download, and no confirmed graduation year. --degree is a caller-supplied index label, not an API fact. identity_state: verified means the row is one source author id, not a hiring check. People who share a display name stay on separate rows.

Default pause: 0.2s for OpenAlex (under the keyless cap of 10 requests/second and the free-key cap of 30) and 1.1s for Semantic Scholar (the introductory API-key limit is 1 request/second). HTTP 429 and 503 responses are retried up to three times, honoring Retry-After.

cd research-talent-scout
export OPENALEX_API_KEY=...   # do not commit
export S2_API_KEY=...         # optional

python3 scripts/fetch_records.py \
  --source openalex \
  --venue NeurIPS \
  --year 2024 \
  --degree phd \
  --output data/openalex.jsonl

python3 scripts/fetch_records.py \
  --source semanticscholar \
  --author-id 1234567 \
  --degree phd \
  --output data/s2.jsonl

python3 scripts/init_index.py --db data/candidates.sqlite
python3 scripts/import_records.py --db data/candidates.sqlite --input data/openalex.jsonl

--author searches by name. Semantic Scholar paper search is relevance-ranked, not a complete census of a venue, and the default cap is 25 works. The first author of a supported main-conference venue is main_first. Workshop, findings, short, and demo are recorded only when the venue string says so. Every other author is unclassified. Equal contribution is not inferred.

The offline sample uses synthetic people (Ada Fixture and Ben Fixture), not a real candidate database:

python3 scripts/import_records.py \
  --db data/candidates.sqlite \
  --input examples/sample-records.jsonl

Demo report

The static sample is examples/demo-report/index.html. It fills the existing report template with the public author order of arXiv:1706.03762, Attention Is All You Need, and the affiliations printed on the paper HTML. Every contact path is “暂未找到”. The sample contains no email address and no mailto link.

Open that file in a browser to see the layout. GitHub renders the file as source. Regenerate it from the template, and check that the committed copy matches:

python3 scripts/render_demo_report.py
python3 scripts/render_demo_report.py --check

The sample stays on main so it can be reviewed in the repository. CI does not publish it to gh-pages.

Demo index

The quick start imports synthetic people from research-talent-scout/tests/fixtures/records.jsonl. Names such as Alice Example are fixtures, not a candidate database.

cd research-talent-scout

python3 scripts/init_index.py --db data/candidates.sqlite
python3 scripts/import_records.py \
  --db data/candidates.sqlite \
  --input tests/fixtures/records.jsonl

python3 scripts/query_candidates.py \
  --db data/candidates.sqlite \
  --graduation-years 2027,2028 \
  --degree-threshold phd=5,masters=3 \
  --strictly-greater \
  --venues ACL,ICML,ICLR,NeurIPS,EMNLP,CVPR,AAAI,ICCV,MLSys,OSDI,EuroSys,ASPLOS,ISCA,SC,PPoPP \
  --categories main_first,co_first,findings,workshop,short,demo \
  --graduation-states confirmed,likely_needs_verification

Verification

Run from the repository root:

bash tests/verify-all.sh

This validates both Skill structures, frontmatter, the original paper-mode contract, the Pages homepage generator, the demo report, and the scout’s deterministic index, import, query, connector-mapping, evidence-state, and rendering tests. Connector tests use checked-in HTTP fixtures and do not touch the network. GitHub Actions runs the same command on pushes and pull requests to main. The frontmatter check is the in-repo standard-library script scripts/validate_skill_frontmatter.py.

GitHub Pages

Paper-mode reports stay off main. When publication is authorized, the Skill publishes:

branch: gh-pages
file:   /index.html
URL:    https://.github.io///
home:   https://.github.io//

For this repository the collection home is https://lofisu.github.io/recruiting-skills/. scripts/build_pages_index.py renders it from assets/pages-index-template.html. If GitHub writes, Pages settings, or deployment checks are unavailable, the Skill returns the local report and does not invent a live URL.

To see the report layout before any Pages publish, open examples/demo-report/index.html. That file is a static sample committed on main, not a report published on gh-pages.

Data boundaries

Only public professional information is collected. The Skills never guess email addresses, log in to restricted platforms, bypass access controls, solve CAPTCHAs, or treat search snippets as verified evidence. Outputs are research leads for human review, not candidate rankings, employment decisions, or automated outreach.

See CONTRIBUTING.md and CODE_OF_CONDUCT.md. Report vulnerabilities in private using SECURITY.md. Licensed under MIT.