Deeptutor

DeepTutor — Academic Advisor Investigation System

Multi-platform skill. This SKILL.md is the entry for Claude Code / agentskills.io. For Codex CLI, OpenCode, OpenClaw, Aider, Cline, Continue, and any other tool following the agents.md spec, the equivalent entry point is the repo-root AGENTS.md. Cursor reads .cursor/rules/deeptutor.mdc. All three documents describe the same workflow; updates should be propagated to AGENTS.md when changing the workflow itself.

Core Principle

Your ceiling = your seniors' ceiling.

Student outcomes are the single most predictive signal for advisor quality. A professor with stellar publications but whose students consistently end up in unclear positions is a red flag. A professor with modest metrics but whose students thrive is gold. Always weight student trajectory evidence above all other dimensions.

Language & Region Detection

Language Rule

The report renderer currently supports Chinese and English. Chinese input produces a fully Chinese report; English input produces a fully English report. If the user requests another language, ask whether an English report is acceptable before starting. Never mix languages within a report except when quoting original sources with a translation.

Region Detection

Determine region from the institution name. This affects which search platforms to use and which evaluation criteria apply.

When uncertain about region, ask the user.

Input Requirements

Minimum input: Professor name + institution name.

If the user hasn't provided these, ask:

1. Career goal (shapes the Goal-Advisor Match scoring dimension)
   • Chinese context: 读博深造 / 考公考编 / 进大厂 / 药企CRO / 进医院 / 纯拿学位
   • International context: Academic career (tenure-track) / Industry R&D / Consulting & Finance / Government & Policy / Startup / Just get degree
2. Risk tolerance: Conservative / Moderate / Aggressive
3. Specific concerns (optional): e.g., "I heard the lab has high turnover"

If the user doesn't provide career goal or risk tolerance, proceed with a balanced evaluation and note that the Goal-Match dimension couldn't be fully scored.

Model Capability Detection & Version Selection

DeepTutor has two investigation modes. The right mode depends on the model running it.

Auto-Detection Rule

Full Version (完整版) — run without asking:

• Claude Opus 4.6+ / Claude Sonnet 4.6+
• GPT-5 / GPT-5-Codex (powering Codex CLI) and equivalent
• Gemini 2.5 Pro / Gemini 3 and equivalent
• Future flagship models of equivalent or higher capability

Prompt user to choose — for all other models (GPT-4o, Gemini Flash, GLM, MiniMax, Claude Haiku, smaller open models, etc.), display:

⚠️ DeepTutor 模式选择 检测到当前模型非旗舰级别。

• 完整版: 10阶段/11维度/18节报告（推荐高端模型）
• 轻量版: 6阶段/7维度/7节报告（Token约完整版40%，可能遗漏部分信息） 请选择：完整版 or 轻量版？

If unsure of the running model's class, default to prompting rather than silently running Full — a Lite report from a weaker model beats a hallucinated Full report.

Lite Version: 6-Phase Workflow

If the user chooses Lite, read references/lite_mode.md for the full specification. Key differences:

• 6 phases (skip co-author network, funding analysis, macro trend deep dive, retirement risk)
• 7 scoring dimensions (merge and drop 4 dimensions, re-weight)
• Simplified Sharp Critique (5-line template instead of 7-question framework)
• 7-section report (instead of 18)

Report Generation (Both Versions)

Both Full and Lite versions should output structured JSON and use scripts/generate_report.py for HTML rendering:

# Model outputs investigation data as JSON → script renders HTML
python scripts/generate_report.py report_data.json -o report.html

This separates investigation (model's job) from rendering (script's job). Even Full version benefits from this — the model focuses on analysis, not wrestling with CSS.

10-Phase Investigation Workflow (Full Version)

Phase 1: Identity Resolution

Establish the professor's verified identity across platforms. This prevents investigating the wrong person (especially common with Chinese names that have many romanization variants).

For all regions:

• Official faculty page (university website)
• Google Scholar profile
• Scopus Author ID / ORCID
• Semantic Scholar

Chinese-specific additions:

• Baidu Scholar (百度学术)
• ResearchGate
• X-MOL faculty profile
• NSFC funded project database (kd.nsfc.cn)
• ScholarMate

International-specific additions:

• DBLP (for CS)
• Web of Science ResearcherID
• Personal/lab website
• GitHub (for computational fields)

Key verification: Cross-reference at least 3 platforms. Confirm institution, department, research area, and photo (if available) all align. For Chinese scholars, generate ALL name romanization variants — see references/publication_search_protocol.md for the template.

Phase 2: Student Trajectory Tracking (THE MOST IMPORTANT PHASE)

This phase implements the "ceiling principle." Track as many current and former students as possible.

How to find students:

• Lab/group website "Members" or "Alumni" page
• Co-authored papers (students are typically first authors)
• University thesis/dissertation databases
• Chinese: CNKI/万方 thesis search, 小木虫 lab discussions
• International: LinkedIn (search "[professor name] lab" or "[university] [department]"), ProQuest Dissertations, university digital repositories

What to track for each student:

Ceiling/Floor analysis:

• High ceiling: Multiple students in tenure-track faculty, top-tier postdocs, or leadership roles in industry
• Mid ceiling: Students in decent positions but not exceptional
• Low ceiling: Multiple verified weak outcomes or frequent attrition; do not infer this from missing public profiles alone
• Information gap: No public student outcomes may mean a new PI, privacy choices, or sparse records. Mark it unresolved and verify directly rather than treating absence as a negative fact.

Phase 3: Publication Analysis

Follow the protocol in references/publication_search_protocol.md EXACTLY. The mandatory rule: always start with a BROAD search (no field keywords), then narrow down.

Search sequence:

1. Broad PubMed/Scopus/Scholar search with name + institution (NO topic keywords)
2. Author ID-anchored search (Scopus ID, ORCID, Semantic Scholar)
3. All name variants from the romanization template
4. Cross-database verification (minimum 3 databases)

Analyze:

• Total output, h-index, i10-index
• Publication trend (increasing/stable/declining)
• Journal quality distribution (top-tier / mid-tier / low-tier)
• Student first-authorship ratio
• Publication gaps (use the 6-step verification checklist before concluding any gap)
• Preprint activity (bioRxiv, arXiv, medRxiv)

Phase 4: Co-Author Network & Advisor Classification

Build a co-author frequency table from the publication record. Classify relationships:

• Internal collaborators (same institution)
• External academic collaborators
• Clinical/industry collaborators
• Student/postdoc co-authors

Advisor Type Classification:

Classify based on: meeting frequency, student authorship patterns, project types (basic vs applied), student independence signals.

Phase 5: Funding Analysis

Chinese institutions → Read references/chinese_academic_system.md:

• NSFC grants (青年/面上/重点/杰青/优青)
• Ministry-level programs (973, National Key R&D)
• Provincial and university internal grants
• Industry/hospital collaboration (横向) funding

International institutions → Read references/international_academic_system.md:

• Government grants (NIH R01/R21, NSF CAREER, ERC Starting/Consolidator/Advanced, EPSRC, DFG, JSPS)
• Foundation grants (HHMI, Wellcome Trust, Gates Foundation)
• Industry funding and consulting
• Startup funds (common for new faculty)

Assess:

• Continuous vs sporadic funding
• Funding trajectory (growing or shrinking)
• Diversity of funding sources
• Whether funding supports student stipends and research

Phase 6: Contextual Intelligence — Social & Reputation Search

This phase uses region-specific platforms to gather student reviews and lab culture signals.

Chinese Mainland Strategy

Search these platforms for: "导师名" + 评价/怎么样/读研/课题组/实验室/push/pua

Also search: university BBS, WeChat public accounts (if accessible), news articles about the professor.

International Strategy

Search these platforms for: "professor name" + "university" + review/advisor/lab/experience/toxic

Also search: department-specific student surveys (some universities publish these), news articles, academic misconduct databases (Retraction Watch).

Hong Kong / Macau / Taiwan Strategy

Combine BOTH Chinese and international platforms, plus:

• PTT (Taiwan: ptt.cc)
• LIHKG (Hong Kong: lihkg.com)
• Dcard (Taiwan/HK student platform)
• 小红书 and 知乎 (many HK/TW students post here)

Phase 6.5: Field Macro Trend Analysis (行业宏观趋势判断)

方向不对，再好的导师也帮不了你。

在完成社会评价搜索后、打分之前，必须对导师所在研究领域进行宏观趋势判断。这不是简单的"hotspot or not"，而是系统性地评估这个领域对学生未来5-10年职业发展的影响。

必须回答的5个核心问题：

1. 生命周期定位：这个领域处于什么阶段？

   • 🌱 萌芽期（Emerging）：新技术/新概念，论文少但增长快，风险高回报高
   • 📈 上升期（Growth）：资金涌入，招聘旺盛，竞争加剧但机会多
   • 📊 成熟期（Mature）：方法论稳定，工业化应用，增量创新为主
   • 📉 衰退期（Declining）：资金缩减，人才外流，被新技术替代
   • ☠️ 夕阳期（Sunset）：几乎无新资金，从业者转行，学生就业极难

2. 资金趋势：近5年该领域的国家级基金（NSFC/NIH/ERC）资助数量和金额是增是减？有没有新的专项计划？

3. 就业市场前景：

   • 学术界：该领域的faculty招聘岗位是否在增加？
   • 工业界：对口企业/岗位有哪些？薪资水平？招聘趋势？
   • 医疗/政府：是否有对口的临床或政策岗位？

4. 技术颠覆风险：该领域是否面临被AI/新技术/新方法论替代的风险？（如：传统组学分析 vs AI驱动的组学，传统药物筛选 vs AI drug discovery）

5. 中国/国际差异：同一个领域在国内和国际的发展阶段可能不同（如：某领域在国内是政策热点但国际已趋于饱和，或反之）

信息来源：

• 领域顶刊的发表量年度趋势（PubMed/Scopus统计）
• 国家基金资助项目数量趋势（NSFC/NIH Reporter）
• 行业报告和市场分析（招聘网站、行业白皮书）
• 领域顶级会议的参会规模变化
• 知名课题组的方向转移信号

输出格式： 给出明确的趋势判断标签（萌芽/上升/成熟/衰退/夕阳）+ 置信度 + 关键证据 + 对学生的具体影响。

Phase 7: Multi-Dimensional Scoring

Read references/advisor_evaluation_framework.md for detailed rubrics.

Chinese context — 11 dimensions:

International context — 11 dimensions:

New dimensions explained:

• Field Macro Trend (D1): Replaces old "Research Direction & Prospects" with a much deeper, structured macro trend analysis (see Phase 6.5). Not just "is it a hotspot" but WHERE in the lifecycle, WHAT the job market looks like, and WHETHER the field faces disruption.
• Advisor Sharp Critique (D10): A synthesized, honest assessment that cuts through diplomatic scoring. See Phase 9.5 for details.
• Retirement & Stability Risk (D11): Evaluates whether the advisor will still be active and funded for the full duration of the student's degree.

Key difference: The Chinese "时间自由度" dimension evaluates freedom for 考公/考编/实习, which is irrelevant for international students. The international "Work-Life Balance" evaluates vacation policy, expected work hours, remote flexibility, and support for career development activities (conferences, internships, courses).

Phase 8: Red / Green Flag Check

Run through the flag checklists in references/advisor_evaluation_framework.md. Region-specific flags:

Universal red flags:

• Multiple documented poor student outcomes or attrition cases; missing public records alone are an information gap, not a red flag
• Extended time-to-degree pattern
• Students leaving mid-program
• No papers in 2+ years
• Funding gaps > 3 years
• Multiple PUA/toxicity reports online
• An official misconduct finding, or a retraction whose verified cause creates a current supervision or lab-safety risk

Chinese-specific red flags:

• 横向 projects with no student benefit
• Only 硕导 but recruiting PhD-track students
• Excessive graduation requirements beyond norms
• No internship permission despite students wanting industry careers

International-specific red flags:

• High postdoc churn rate
• Lab members rarely listed as first/corresponding author
• No conference travel support
• Visa sponsorship issues for international students
• Advisor takes credit for student work (scooping)
• "Revolving door" lab (many short-tenure members)
• Glassdoor/Reddit reports of toxic culture

Universal green flags:

• Multiple student first-author papers in good journals
• Clear, positive student outcomes
• Recent promotion or awards
• Conference support for students
• Reasonable stipends
• Positive online reviews from current/former students

Phase 9: Advisor Sharp Critique (导师锐评)

不要让外交辞令害了学生。学生需要的不是3.8分还是4.1分的区别，而是"这个人到底能不能选"的直觉判断。

这个阶段是整个评估的灵魂。在完成所有数据收集和机械化打分后，用以下框架对导师进行一次不留情面的直觉评估。

锐评必须回答的7个问题：

1. 一句话判决：如果你的亲弟弟/亲妹妹问你能不能选这个导师，你会说什么？（不是写给学术委员会的，是写给家人的）

2. 导师的"人设"vs现实：

   • 导师对外展示的形象是什么？（官网简介、招生宣传、公开讲话）
   • 数据和学生评价反映的现实是什么？
   • 两者之间有多大差距？差距越大越危险。

3. 最大的隐藏风险：导师不会主动告诉你、但你入组后一定会遇到的问题是什么？（基于学生评价、出组率、发表模式推断）

4. 最被低估的优点：导师身上被分数系统低估的、真正有价值的特质是什么？

5. 5年后预测：根据官方任期/招生安排、职称、资金、发表趋势、领域走向——5年后这个实验室会是什么状态？上升、稳定、还是衰退？

6. 替代方案建议：如果不选这个导师，在同一领域/同一学校，还有什么替代选择值得考虑？（基于合作者网络和同院系信息推断）

7. Deal-Breaker检查：是否存在以下任何一个"一票否决"条件？

   • 多条独立的PUA/toxicity投诉（不是一条可能是个人恩怨，多条就是系统性问题）
   • 导师3年内即将退休但没有明确的接班安排
   • 近3年完全无经费且无新论文
   • 多名学生中途退组/延期毕业的明确证据
   • 有官方结论的严重学术不端，或有文件记录、会造成当前指导/实验室安全风险的报复行为；撤稿本身不等于不端，必须核验原因与处理结果
   • 如果触发任何一条，无论其他维度分数多高，总评必须标注为"⚠️ 存在一票否决风险"

锐评的评分标准：

锐评的写作风格：

• 说人话，不说学术套话
• 用具体事实支撑判断，不空谈
• 敢于给出明确的"推荐/不推荐"结论，不骑墙
• 如果信息不足无法判断，直说"信息不足，无法给出可靠的锐评"，不要硬编

Phase 9.5: Retirement & Stability Risk Assessment

评估导师在学生就读期间是否会保持稳定。

检查项：

• 官方公开的任期、退休或招生安排；不要根据照片、姓名、性别或其他代理信息猜测年龄
• 当前国家/地区法律与学校政策（必须查当期官方来源，不使用固定年龄的跨地区假设）
• Tenure status（国际）：pre-tenure PI有被deny tenure导致实验室关闭的风险
• 经费连续性：当前经费何时到期？是否有续期迹象？
• 是否有实验室搬迁/跳槽迹象？（关注近期的职位变动、多个affiliation）
• 活跃度与招生承诺：近期是否持续招生、指导学生、申请经费和发表成果；不要据此推断健康状况

Phase 10: Report Generation

Output all investigation data as structured JSON, then render via scripts/generate_report.py:

python scripts/generate_report.py report_data.json -o "教授名_机构.html"

The JSON schema and 18-section report structure are defined in references/report_template.md. Key rules: use the selected Chinese or English language consistently, cite every claim, and keep 锐评 in the top 3 sections.

Parallel Search Strategy

Launch searches in parallel batches to maximize efficiency:

• Batch 1: Faculty page + Scholar profile + Lab website + Thesis DB
• Batch 2: PubMed broad (NO keywords) + Scopus/OpenAlex + Name variants + Preprints
• Batch 3: Social platforms + News + Funding DBs + Retraction Watch
• Batch 4: Cross-validate counts + Verify student outcomes + Fill gaps

Quality Rules

1. Every claim needs a source. No unsourced assertions.
2. Distinguish fact from inference. Mark speculative conclusions explicitly.
3. Cross-validate metrics. Use ≥3 databases for publication counts.
4. Weight recent evidence. Last 5 years matter more than career totals.
5. No fabrication. If information is unavailable, say so — don't guess.
6. Be balanced. Report both strengths and weaknesses.
7. Score vs peers. Compare against others at the same institution and rank.
8. Student signals > publication metrics. Always.
9. PUA/toxicity evidence is critical. Don't downplay concerning signals.
10. Publication gap verification. Complete the 6-step checklist before concluding any gap.

Privacy & Responsible Use

1. Use public, professionally relevant information only. Do not seek or publish home addresses, private phone numbers, personal accounts, family details, or data behind access controls. search_social.py sends query terms to Bing and may fall back to Sogou, so never include private or sensitive information in a query.
2. Prefer official institutional contact details. Include a student's identity only when it is already public in a professional context and necessary for a sourced trajectory claim.
3. Treat search snippets and anonymous posts as leads, not verified facts. Corroborate serious allegations with independent evidence; label unresolved claims as allegations and avoid reproducing handles or identifying details.
4. Do not infer age, health, ethnicity, gender, family status, or other sensitive traits from names, photos, publication pace, or proxies.
5. Keep fact, allegation, and inference visibly separate. A report supports the user's due diligence; it is not a misconduct finding or a substitute for speaking with current and former students.

Comparative Mode

When comparing multiple advisors: investigate each independently, generate individual reports, then add a comparison card with side-by-side scores, composite comparison, and trade-off analysis.

Integration with Other Skills

Leverage these skills when available:

• pubmed-database, openalex-database — Publication searches
• deep-research, exa-search — Web research and social platform mining
• biorxiv-database, arxiv-database — Preprint searches
• scientific-visualization, matplotlib — Charts in report
• literature-review — Systematic publication analysis
• citation-management — Reference verification
• scripts/robust_fetch.py — Public-network-only stdlib fetch with DNS/IP pinning; optional curl/browser interfaces fail closed (adapted from Web-Rooter, MIT; see THIRD_PARTY_NOTICES.md)
• scripts/search_social.py — Chinese social platform search (知乎/小红书/小木虫/贴吧/保研论坛/考研帮)

Chinese Website Fallback (zero dependencies, details in references/web_rooter_integration.md)

python scripts/robust_fetch.py "<URL>"                                     # pinned stdlib HTTP(S) only
# --js is intentionally disabled because browser networking cannot preserve the pinned-IP boundary
python scripts/search_social.py "导师名 大学名" --platforms zhihu,xiaohongshu,emuch  # social search

• URL freshness: if 404, re-search via WebSearch("site:<domain> 教授姓名")
• If wr available: prefer wr html/wr social (see references/web_rooter_integration.md)