A Claude Code / Codex skill that turns a short conversation into a researched, interactive marimo dashboard for planning PhD applications: find programs, surface fitting advisors (PIs), compare stipends and application rules, and decide where to apply.
It works for any field, any region. You tell it what you care about; it researches in parallel when the host supports workflow/subagent tools, then hands you a dashboard you can filter, annotate, and rank.
No personal data is baked in. Every run is driven entirely by your own intake answers.
A single-page marimo app with:
- KPIs and a top-fit picks grid
- Program browser: filter by region / fit / stipend / application-hedge; aligned table rows; click x on any row to hide it and restore later
- Program detail cards: stipend with verification basis, cohort, deadlines, application restrictions, career outcomes, international-student notes
- PI browser: filter by category / wet-dry / region / min-h-index / keyword; per-row hide; editable notes per advisor that persist across restarts; Google Scholar links
- Weighted priority ranker: 5 sliders plus interest picks, live re-ranking
- Cost-of-living-adjusted real stipend chart
- Statistics: Spearman correlation, Kruskal-Wallis, Pareto frontier, K-means clustering plus PCA, with plain-language, stats-based recommendations
Your notes and hidden lists are saved next to the data, so they survive restarts.
- Claude Code or Codex with local Skills enabled
- Python 3.9+ with:
pip install marimo pandas matplotlib scipy scikit-learn
- Web access for the research step
Claude Code can use the bundled assets/research_workflow.js Workflow template. Codex can use
the same SKILL.md, reference/, and assets/ package; when a Workflow tool is unavailable,
the skill performs equivalent web research and writes the same _wf_result.json schema before
building the dashboard.
Clone into your Claude Code skills directory:
mkdir -p ~/.claude/skills
git clone https://github.com/SihengTao/phd-application-planner \
~/.claude/skills/phd-application-plannerClaude Code auto-discovers skills in ~/.claude/skills/.
Clone into your Codex skills directory:
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
git clone https://github.com/SihengTao/phd-application-planner \
"${CODEX_HOME:-$HOME/.codex}/skills/phd-application-planner"Codex discovers local skills from ${CODEX_HOME:-$HOME/.codex}/skills/. The included
agents/openai.yaml provides Codex-facing display metadata; the root SKILL.md is the
shared execution guide.
Ask Claude Code or Codex, in any project:
"Help me find PhD programs in and build me a dashboard."
You can also invoke it explicitly:
- Claude Code:
/phd-application-planner - Codex:
Use $phd-application-planner to help me build a PhD application dashboard.
-
Interactive intake. It asks what you care about: research field and subfields, target regions/countries, minimum stipend and currency, how many programs and advisors per program, PI preferences, rising-star bias, application constraints, and lifestyle/outcome concerns. This is usually 3-5 quick question rounds.
-
Config. It writes a
_config.jsonwith dashboard branding, regions, stipend floor, and an auto-generated interest-tag taxonomy for your field. -
Research. In Claude Code, it can run the bundled Workflow template for parallel research. In Codex, or in any host without Workflow, it performs the same program / PI / outcome / verification research path and saves the final object as
_wf_result.json. Stipends and PIs should be verified against official pages and Google Scholar; "not found" stays honest. -
Build data. It assembles the results into the dashboard's data files.
-
Launch. It copies the marimo app next to the data and starts it, opening it in your browser.
- Browse programs and advisors; hide the ones you're not interested in.
- Write notes on each advisor as you meet/read about them; notes persist.
- Use the priority ranker sliders to weight fit / stipend / city / advisor-density / topics.
- Read the statistics section for data-driven strategy tips.
Re-running the research refreshes the program/PI data but keeps your notes and hidden lists.
phd-application-planner/
├── SKILL.md # shared Claude Code / Codex skill instructions
├── agents/
│ └── openai.yaml # Codex skill display metadata
├── assets/
│ ├── dashboard_template.py # the marimo app
│ ├── research_workflow.js # Claude Workflow research template
│ ├── build_data.py # research output -> dashboard data files
│ └── launch.py # detached marimo launcher
├── reference/
│ ├── intake.md # intake question bank
│ ├── schema.md # data schema
│ └── honesty.md # verification / no-fabrication rules
├── README.md
└── LICENSE
- Regions are arbitrary: US, EU, UK, Canada, Singapore, China, etc. Each gets a color and
short label in
_config.json; the whole dashboard adapts. - Seed programs: if you already have a target list, it is passed straight into the research step instead of auto-discovery.
- Language: the dashboard UI labels ship in Chinese; translate strings in
dashboard_template.pyif you prefer another language. Titles/subtitles are configurable per run via_config.json.
Stipends, deadlines, cohort sizes, application rules, and h-indexes are web-research
snapshots. They can be wrong or out of date. Always confirm the numbers that matter on the
program's official pages before deciding. See reference/honesty.md for the verification rules
the skill follows.
这是一个同时适用于 Claude Code / Codex 的 skill。它会把一段简短对话转成一个经过调研的 交互式 marimo 仪表盘，用来规划博士申请：寻找项目、筛选匹配的导师 （PI）、比较 stipend 和申请规则，并帮助你决定申请哪些项目。
它适用于任何领域、任何地区。你只需要说明自己在意什么；如果运行环境支持 workflow 或 subagent，它会并行调研，然后交付一个可以筛选、做笔记、隐藏、排序的仪表盘。
仓库里不内置任何个人信息。每次运行都完全由你当次回答的 intake 信息驱动。
一个单页 marimo 应用，包含：
- 关键指标和最匹配项目列表
- 项目浏览器：按地区、匹配度、stipend、申请 hedging 过滤；可以隐藏/恢复单行
- 项目详情卡：stipend 及验证依据、cohort、deadline、申请限制、career outcome、 国际学生相关备注
- PI 浏览器：按类别、wet/dry、地区、最低 h-index、关键词过滤；可以隐藏 PI；可以给每个 PI 写持久化笔记；带 Google Scholar 链接
- 加权优先级排序器：5 个滑块加兴趣方向选择，实时重新排序
- 生活成本调整后的实际 stipend 图
- 统计分析：Spearman 相关、Kruskal-Wallis、Pareto frontier、K-means clustering 和 PCA， 并给出易读的策略建议
笔记和隐藏列表会保存在数据旁边，重启后仍然保留。
- Claude Code 或 开启本地 Skills 的 Codex
- Python 3.9+，并安装：
pip install marimo pandas matplotlib scipy scikit-learn
- 调研阶段需要联网
Claude Code 可以使用仓库里的 assets/research_workflow.js Workflow 模板。Codex 可以使用同一个
SKILL.md、reference/ 和 assets/ 包；如果没有 Workflow 工具，skill 会执行等价的网页调研，
写出同样结构的 _wf_result.json，再继续构建仪表盘。
克隆到 Claude Code 的 skills 目录：
mkdir -p ~/.claude/skills
git clone https://github.com/SihengTao/phd-application-planner \
~/.claude/skills/phd-application-plannerClaude Code 会自动发现 ~/.claude/skills/ 下的 skills。
克隆到 Codex 的 skills 目录：
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
git clone https://github.com/SihengTao/phd-application-planner \
"${CODEX_HOME:-$HOME/.codex}/skills/phd-application-planner"Codex 会从 ${CODEX_HOME:-$HOME/.codex}/skills/ 发现本地 skills。仓库中的
agents/openai.yaml 提供 Codex 侧的展示元数据；根目录的 SKILL.md 是 Claude Code 和 Codex
共用的执行说明。
在 Claude Code 或 Codex 的任意项目里直接说：
“Help me find PhD programs in and build me a dashboard.”
也可以显式调用：
- Claude Code：
/phd-application-planner - Codex：
Use $phd-application-planner to help me build a PhD application dashboard.
-
交互式 intake。 它会询问你的研究领域/子方向、目标地区、最低 stipend 和币种、想调研的 项目数量和每个项目的 PI 数量、PI 偏好、是否偏好 rising-star、申请机制限制，以及城市/就业 等考虑。通常是 3-5 轮简短问题。
-
写配置。 它会生成
_config.json，包含仪表盘标题、地区、stipend floor，以及根据你领域 自动生成的兴趣标签体系。 -
调研。 在 Claude Code 里，它可以运行内置 Workflow 模板进行并行调研。在 Codex 或没有 Workflow 的环境里，它会执行同样的项目 / PI / outcome / 验证流程，并把最终结果保存为
_wf_result.json。Stipend 和 PI 应该根据官方页面和 Google Scholar 验证；找不到就诚实标注 “not found”。 -
构建数据。 它会把调研结果整理成仪表盘读取的数据文件。
-
启动。 它会把 marimo app 复制到数据旁边并启动，在浏览器中打开。
- 浏览项目和 PI；隐藏不感兴趣的条目。
- 给每个 PI 写笔记，之后重启也会保留。
- 用优先级排序器调整 fit / stipend / city / advisor density / topics 的权重。
- 阅读统计部分，辅助形成申请策略。
重新调研会刷新项目和 PI 数据，但保留你的笔记和隐藏列表。
phd-application-planner/
├── SKILL.md # Claude Code / Codex 共用 skill 说明
├── agents/
│ └── openai.yaml # Codex skill 展示元数据
├── assets/
│ ├── dashboard_template.py # marimo 应用
│ ├── research_workflow.js # Claude Workflow 调研模板
│ ├── build_data.py # 调研输出 -> 仪表盘数据文件
│ └── launch.py # 后台 marimo 启动器
├── reference/
│ ├── intake.md # intake 问题库
│ ├── schema.md # 数据 schema
│ └── honesty.md # 验证 / 不编造规则
├── README.md
└── LICENSE
- 地区是任意的：美国、欧盟、英国、加拿大、新加坡、中国等都可以。每个地区在
_config.json里有颜色和短标签，仪表盘会自动适配。 - Seed programs：如果你已经有目标项目列表，会直接传入调研步骤，而不是从头自动发现。
- 语言：仪表盘 UI 默认是中文；如果想换语言，可以修改
dashboard_template.py中的字符串。 标题和副标题也可以在每次运行的_config.json里配置。
Stipend、deadline、cohort size、申请规则和 h-index 都是网页调研快照，可能错误或过期。决定前
务必到项目官方页面确认关键数字。reference/honesty.md 里记录了这个 skill 使用的验证规则。