daily-job-picks

name

daily-job-picks

description

Find and curate daily high-quality foreign-company China roles and overseas remote roles for China-based applicants, including public multi-category daily roundups and targeted searches for a specific role profile. Use when the user asks for daily job picks, China-applicable foreign company jobs, APAC/Asia remote jobs, overseas remote jobs compatible with China time zones, group-friendly job newsletters, curated job leads with deduplication by date, or parallel multi-agent job-search workflows.

Daily Job Picks

Purpose

Curate a daily list of high-quality jobs that China-based applicants can realistically apply to. Support both public multi-category roundups for friends from different backgrounds and targeted searches for a specific role profile. Prioritize quality over quantity, avoid repeats from previous daily picks, and output date-stamped Markdown records in the current project.

This skill is source-driven. Editable source groups, role profiles, screening rules, link rules, and output defaults live in sources/job-search-config.toml. SKILL.md defines the workflow and non-negotiable safety constraints; the TOML file defines the normal search and screening inputs.

Do not assume a personal default job profile. In targeted mode, the user must provide role keywords, industry, seniority, or other search parameters, or match an enabled role profile from sources/job-search-config.toml. In public roundup mode, use enabled role profiles from the TOML source as the coverage pool.

Files

• Use JOB_PICKS_OUTPUT_ROOT when it is set. Otherwise, default to <current project>/job-picks/.
• Write public daily roundups to <output_root>/YYYY-MM-DD.md.
• Write targeted runs to <output_root>/YYYY-MM-DD-<topic>.md, where <topic> is a short hyphen-case or Chinese-safe label from the requested role or industry.
• Create <output_root>/ if it does not exist.
• New public roundup files must use the top-level title # YYYY-MM-DD 外企/海外远程岗位精选.
• New targeted files must use a distinct top-level title, preferably # YYYY-MM-DD <目标岗位/方向> 岗位专选, so multiple runs on the same day are distinguishable in file lists and previews.
• If writing to an existing date file, append a new ## <run label> section under the existing date file instead of repeating the top-level # YYYY-MM-DD ... title. Use labels such as 默认公共精选, 低英文门槛精选, or the user's requested role/topic.
• Maintain <output_root>/seen-jobs.tsv as a lightweight deduplication index with columns: date, title, company, url, job_direction, source.
• Before selecting jobs, scan both seen-jobs.tsv and existing Markdown files in <output_root>/ and exclude any job whose company plus title or job URL already appeared.
• If the user asks for a different output path, follow the user path and still use date-named Markdown files unless told otherwise.
• Treat sources/job-search-config.toml as the editable source of truth for normal source lists, role profiles, screening preferences, and output defaults.
• If JOB_PICKS_AUDIENCE_PREFERENCES is set, read that file as the project-level audience preference override. Otherwise, if <current project>/config/job-picks-audience-preferences.md exists, read it before searching. It is intentionally human-readable Markdown for local group demand, not a replacement for the skill's TOML source list or hard safety rules.

Source Configuration

• Always read and validate sources/job-search-config.toml before searching.
• Use scripts/validate_source_config.py to validate the TOML source. Stop and report the configuration errors if validation fails.
• Configuration precedence is: user input for the current run > project-level audience preferences from JOB_PICKS_AUDIENCE_PREFERENCES or <current project>/config/job-picks-audience-preferences.md > sources/job-search-config.toml > non-negotiable safety constraints in this SKILL.md.
• User input may override target roles, industry, seniority, language constraints, excluded industries, and count for the current run. Do not write those temporary overrides back to the TOML file unless the user explicitly asks to update the source.
• Project-level audience preferences may override public roundup emphasis, target count, role-direction priority, work-mode priority, English-level tendency, seniority tendency, and applicant-barrier tendency. They must not override hard safety rules, link rules, bad-link exclusions, or explicit company/domain exclusions.
• Public roundup mode must use enabled role_profiles from the TOML source for coverage. Do not require every profile every day; use them to build the candidate pool and keep quality high.
• Targeted mode must first try to match the user's requested role/industry against enabled role_profiles by id, label, keywords, or directions. If no profile matches, use the user's words as temporary role keywords for that run.
• Use enabled source_groups from the TOML source as the primary search lanes. Do not use disabled source groups.
• trust_level = "A" sources are preferred for final links. trust_level = "B" sources can be final links only when they are public, specific, active job pages. trust_level = "C" sources are discovery-only when avoid_as_final_link = true; find a canonical employer or ATS URL before selecting the job.
• Apply TOML screening_rules.excluded_companies and screening_rules.excluded_domains before finalist review. These exclusions are especially useful for companies or domains that are technically active but unsuitable for the audience.
• references/search-and-screening.md is explanatory reference material. Use it to interpret edge cases, but do not treat it as the primary editable source list.

Scripts

Use bundled scripts for deterministic bookkeeping and output. Run them from the skill folder or pass absolute paths.

• scripts/seen_jobs.py: create/read/update seen-jobs.tsv, normalize URLs, and check duplicates against TSV plus prior Markdown files.
• scripts/bad_links.py: create/read/update bad-links.tsv and avoid links or company-title pairs that users reported as broken, closed, paywalled, or not publicly useful.
• scripts/ats_extract.py: extract basic public fields from common ATS/job-detail pages to reduce title/company/link mistakes. Use this as an aid, not as the final truth.
• scripts/link_check.py: run basic reader-facing URL checks. Use it as a first-pass filter; still open finalist links directly before final output.
• scripts/format_daily_picks.py: validate final structured jobs and render the required Markdown format.
• scripts/validate_report.py: validate the rendered Markdown report for required fields, link format, count consistency, duplicate sections, and forbidden internal wording.
• scripts/validate_source_config.py: validate sources/job-search-config.toml using Python standard-library tomllib; requires Python 3.11+ and no third-party package.
• scripts/resolve_search_plan.py: resolve the user's requested mode/role plus TOML source into matched profiles, keywords, source lanes, and search queries. Use this to reduce interpretation drift.

These scripts do not decide whether a job is good or China-applicable. The agent still owns source search, JD interpretation, risk judgment, time-zone judgment, and final selection.

Modes

• Public roundup mode: Use when the user asks for a daily selection for friends, a group, a newsletter, or people from multiple industries. Cover several job directions so readers can self-pick.
• Targeted mode: Use when the user specifies a role, industry, seniority, or personal profile. Optimize for that request instead of broad coverage.

Workflow

1. Confirm the run date using the current date from the environment. Use that date in the output filename and final note.
2. Identify the mode. If the user did not specify a mode, infer public roundup mode for group/friend/newsletter wording and targeted mode for role/profile wording.
3. Run scripts/validate_source_config.py --summary. If it fails, stop before searching and report the TOML configuration errors.
4. Load sources/job-search-config.toml and parse enabled source_groups, enabled role_profiles, screening_rules, link_rules, and output_defaults. Also load references/search-and-screening.md only as explanatory reference for edge cases.
5. If JOB_PICKS_AUDIENCE_PREFERENCES is set, read that file before resolving public roundup search emphasis. Otherwise, if <current project>/config/job-picks-audience-preferences.md exists, read it:
   • Use 本期总数 as the preferred public roundup target unless the user gives a count.
   • Use 岗位方向优先级 to decide which role families get more search effort and finalist slots.
   • Use 工作方式优先级, 英文要求倾向, 经验阶段倾向, and 申请门槛倾向 as tie-breakers when choosing between otherwise similar jobs.
   • Use 明确排除 as project-specific screening guidance, but never as permission to weaken hard rules.
   • Keep the file human-facing; do not require TOML/YAML syntax from the user.
6. Parse the user's current parameters: role, industry, seniority, skill constraints, language requirements, excluded industries, requested count, and whether broad public coverage is desired.
7. Run scripts/resolve_search_plan.py --mode ... --role ... --industry ... to resolve the search plan from user input plus TOML:
   • Public roundup mode: use enabled TOML role profiles as the coverage pool and use TOML count defaults unless the user specified a count.
   • Targeted mode: match user input to enabled TOML role profiles. If no profile matches, use the user input as temporary keywords for this run.
   • Build search queries from enabled TOML source groups and their search_templates, replacing {role} with role/profile keywords.
   • Apply excluded_companies and excluded_domains from the resolved plan before spending review time on candidates.
8. Run scripts/seen_jobs.py ensure and scripts/bad_links.py ensure, then run scripts/seen_jobs.py snapshot --format json plus scripts/bad_links.py snapshot. Keep a compact dedupe and bad-link snapshot for this run.
9. If subagents are available and the requested search is broad, use the dispatcher workflow in references/multi-agent-workflow.md: the main agent assigns distinct source/category lanes to child agents, then owns deduplication, final screening, writing, and final link validation. If subagents are unavailable or the request is narrow, run the same lanes sequentially.
10. Search live job sources. Because job listings change frequently, always browse the web for current listings. Use enabled TOML source groups as the search lanes and do not use disabled source groups.
11. Build a candidate pool larger than the requested count. For each promising candidate, run scripts/seen_jobs.py check --title ... --company ... --url ... and scripts/bad_links.py check --title ... --company ... --url ...; remove duplicates and previously reported bad links before spending more review time.
12. Reject jobs that fail the hard rules:

• Overseas jobs must be remote.
• Do not select roles with China time-zone incompatibility over 5 hours unless the user explicitly requested them.
• Do not select generic overseas AI Trainer, data annotation, rater, evaluator, or language-model training contractor roles for Mandarin, Simplified Chinese, or Chinese unless the job page explicitly lists mainland China/China as an accepted work location or hiring jurisdiction. Chinese-language ability alone is not evidence that China-based applicants can apply; many such roles target overseas Chinese speakers.
• Do not select obvious scams, gray-market roles, high-risk crypto projects, gambling, adult industry, paid-to-apply jobs, brush-order work, pure pyramid/referral schemes, or vague high-pay roles with unclear company identity.
• Do not select platform homepages, search pages, company job-board landing pages, category pages, expired pages, pages that no longer show the selected role, or pages that open but no longer expose the job description and application path to a normal reader.

13. Apply TOML screening_rules, project-level audience preferences, and link_rules as practical screening inputs, but never use them to weaken the hard rules above.
14. For Greenhouse, Lever, Ashby, Workable, SmartRecruiters, and company career URLs, run scripts/ats_extract.py <url> when shell network access is available. Use extracted title/company/location as a consistency check; if it disagrees with the candidate, open the page and resolve the mismatch before proceeding.
15. MANDATORY link check — do not skip, do not proceed to step 16 until complete. Run scripts/link_check.py --url <url> --title <title> --company <company> for every finalist URL one by one. Any URL that returns "ok_basic": false, an HTTP error, a bad-page marker, missing selected-role details, or no reader-visible application path must be dropped and replaced before continuing. After the script passes, also open each finalist URL directly and run the final reader-usability pass from references/search-and-screening.md; replace or reject any job whose link cannot be verified. For SmartRecruiters and similar ATS pages, do not trust stale page titles, cached snippets, or metadata alone: the currently opened page must show the selected job title, company, job description, and an apply/interested action to a normal reader. Record every dropped link via scripts/bad_links.py append before searching for a replacement. This verification is internal; do not include a 链接核验 field or mention scraping, crawling, rendering, parser behavior, ATS quirks, or verification mechanics in the public output.
16. Classify and summarize each selected job into the structured JSON schema below. Only jobs that passed step 15 may appear here.
17. Run scripts/format_daily_picks.py --input <final-jobs.json> ... to validate required fields and render Markdown. For targeted runs, pass --mode 定向精选 --target "<用户请求的目标岗位/方向>"; the renderer will automatically produce a 岗位专选 title unless --title is provided. Fix any validation errors before writing final output.
18. Save or append the rendered Markdown to the appropriate file, then run scripts/validate_report.py <report.md> --check-links whenever shell network access is available. Fix validation errors before responding. If the result includes bad_link_candidates, record each failed URL with scripts/bad_links.py append before replacing it. If shell network is unavailable, run scripts/link_check.py on every final URL separately as soon as access is available; do not rely on Markdown-only validation.
19. For each accepted job, run scripts/seen_jobs.py append --date ... --title ... --company ... --url ... --job-direction ... --source .... Also provide the same content in the response unless the user only asked to save it.

If the user reports a broken, closed, login-gated, paywalled, wrong-job, or unavailable link from a previous report, run scripts/bad_links.py append --date ... --url ... --title ... --company ... --reason ... before finding a replacement. Treat that report as decisive for future public usefulness.

Quality Bar

Prefer 6-10 strong jobs for public roundup mode and 5-8 strong jobs for targeted mode. Return fewer if not enough high-quality, current, non-duplicate jobs meet all rules. Do not pad with weak matches.

Use direct evidence from the job page for location, remote status, company, time-zone feasibility, role requirements, and application path. If a field is unclear but the role is otherwise promising, label it as 中国可投待确认 and state what the applicant should confirm, using applicant-facing wording such as 投递前确认中国大陆雇佣/合同形式. Do not expose internal collection or verification details such as 抓取, 爬取, 结构化字段, 无登录环境, 页面渲染, ATS, 解析, 检索结果, or 不同环境显示不完全.

For AI Trainer, data annotation, rater, evaluator, and language-model training roles, require stronger China eligibility evidence than for normal remote jobs. A title like Mandarin Chinese AI Trainer, Simplified Chinese Evaluator, or Chinese Data Annotator only proves language demand, not mainland China eligibility. Reject these roles unless the page explicitly supports applicants working from mainland China/China, or the user asks for overseas Chinese applicants instead of China-based applicants.

Required Output Format

Use Chinese for the analysis, keep the original English job title and company name when applicable, and include a clickable direct job link. Split 岗位归类 from 岗位方向; do not overload one field with both eligibility category and role category.

Output style rules:

• Do not include a 链接核验 field in final Markdown files or user-facing responses. Link verification is mandatory but internal.
• The 链接 field must be a Markdown link with the visible text 直达链接, for example 链接：[直达链接](https://example.com/job/123). Do not display the raw URL as the link text.
• In the header/summary metadata, 数量 must report the actual number of selected jobs, not the requested target range. For example, if the target was 5-8 and 5 jobs were selected, write 数量：5 个; do not write 数量目标：5-8 个.
• Do not mention scraping/crawling/search mechanics, parser behavior, login-less environments, structured fields, link verification, or page-rendering limitations in 注意事项 or any public-facing text.
• If eligibility or employment structure is unclear, describe only the applicant-facing uncertainty, such as 投递前确认是否支持中国大陆远程签约、税务和合同形式.
• Infer 岗位方向 from the job title and JD responsibilities, not just from broad skill defaults. Use 其他 only when no more specific label fits after reading the JD.
• Write 申请门槛 using only the application_barrier_note text in applicant-facing wording, for example 申请门槛：需要基础客服或 SaaS 支持经验，但不是重资历岗. Do not render 低 / 中 / 高 in final Markdown or user-facing responses.
• Write 中国可投把握 as 高 / 中 / 待确认 + 一句简短说明 so readers know the evidence behind the judgment, for example 中国可投把握：高，岗位直接面向中国团队或中国本地招聘.

Top-level title rules:

• Public roundup mode: # YYYY-MM-DD 外企/海外远程岗位精选
• Targeted mode: # YYYY-MM-DD <目标岗位/方向> 岗位专选

# YYYY-MM-DD 外企/海外远程岗位精选

筛选参数：
- 模式：公共精选 / 定向精选
- 目标岗位：
- 行业/方向：
- 数量：

### 1. 岗位名称：
公司 / 平台：
岗位归类：外企中国岗位 / 外企 APAC 岗位 / 海外远程岗位 / 中国可投待确认
岗位方向：客服 / 运营 / 内容 / 本地化 / 销售支持 / 技术 / 技术支持 / QA / 数据 / AI Trainer / 产品 / 项目管理 / 需求分析 / 系统分析 / 实施 / 解决方案 / HR / 供应链 / 销售 / 合同工 / 兼职 / 其他
工作方式：中国本地办公 / 混合办公 / 全球远程 / APAC 远程 / 中国可投待确认
经验要求：入门 / 1-3 年 / 3-5 年 / 高级 / 不明确
语言要求：中文 / 英文 / 双语 / 其他 / 不明确
申请门槛：直接使用 application_barrier_note 的一句简短说明
中国可投把握：高 / 中 / 待确认 + 一句简短说明
时差判断：
适合谁：
注意事项：
链接：[直达链接](URL)

### 2. 岗位名称：
公司 / 平台：
岗位归类：
岗位方向：
工作方式：
经验要求：
语言要求：
申请门槛：
中国可投把握：
时差判断：
适合谁：
注意事项：
链接：[直达链接](URL)

这些岗位的筛选时间是 YYYY-MM-DD，申请前仍需以岗位页面最新信息为准。

If a near-match replaces an unavailable category, state the replacement reason in 注意事项 using applicant-facing wording. Keep 中国可投把握 as a practical label, not a mathematical score. 申请门槛 should stay as a concise applicant-facing note from application_barrier_note, not a normalized level and not a dump of JD details. Never output a job whose direct link was not verified unless the user explicitly allows unverified leads; if allowed, still do not add 链接核验, and instead state the practical limitation in 注意事项, for example 链接状态需投递前再次确认.

Final Jobs JSON Schema

Before rendering Markdown, create a temporary JSON file containing the final jobs as an array. Use these exact keys so scripts/format_daily_picks.py can validate and render the output:

[
  {
    "title": "Customer Support Specialist",
    "company_platform": "Company / Greenhouse",
    "company": "Company",
    "job_group": "海外远程岗位",
    "job_direction": "客服",
    "work_mode": "APAC 远程",
    "experience": "1-3 年",
    "language": "英文",
    "application_barrier": "中",
    "application_barrier_note": "需要基础客服或 SaaS 支持经验，但不是重资历岗",
    "china_applicability": "中",
    "china_applicability_note": "面向 APAC 远程，但投递前确认是否支持中国大陆签约",
    "timezone_judgment": "面向 APAC，和北京时间协作可行",
    "best_for": "适合英语较好、有客服或 SaaS 支持经验的人",
    "notes": "投递前确认是否支持中国大陆远程签约、税务和合同形式",
    "url": "https://example.com/jobs/123",
    "source": "Greenhouse"
  }
]

company and source are used for seen_jobs.py append; the rendered Markdown uses company_platform. application_barrier and china_applicability keep normalized internal labels for screening and tie-break decisions. In rendered Markdown, 申请门槛 must use only application_barrier_note, while 中国可投把握 continues to render the base label plus china_applicability_note.