ai-job-search：35k Star，一位地球物理学家用 Claude Code 把求职工程化，69 投 20 面 1 offer

by Mycelium Protocol

---

GitHub：MadsLorentzen/ai-job-search ⭐ 35,452 | Forks 12,170 | Python | MIT
创建：2026-03-18 | 最近更新：2026-08-26

---

先说一个细节

作者 Mads Lorentzen，地球物理学家，2025 年底被裁员。他没有更新 LinkedIn 等通知，而是花时间造了这套工具，用它跑了自己的求职流程：

• 69 份定制申请
• 20 次初试
• 1 份 offer
• 2026 年 6 月入职，职位：AI 工程师

他在每家公司都主动告知自己用了 AI 辅助求职，结果没有一次成为劣势——反而几乎每次都引发了技术对话。随后他把这套工具开源，现在有 35,452 个 Star。

---

优雅的递归

用户提炼得很准：程序员们终于把”用 AI 替代自己工作”的能力，用在了”找到下一份被 AI 替代的工作”上面。

这不是反讽，是现实。求职本身是一个极度重复、高度标准化的信息处理流程：读职位描述、评估匹配度、改简历、写 Cover Letter、做面试准备……这些工作 AI 做得比人耐心、比人一致，而且不会在第 50 封信的时候开始偷懒。

---

核心工作流

/setup        /scrape           /apply <url>
  |               |                  |
  v               v                  v
建立档案      搜索职位           评估匹配度
（简历/      多平台去重         评分 + 建议
LinkedIn/    按匹配度排序
面试）            |                  |
                  v                  v
              选中职位          起草简历 + Cover Letter
              → /apply          （LaTeX，定制化）
                                     |
                                     v
                                审稿 Agent 审核
                                → 修订 → 最终输出

/setup：建立候选人档案

三条路径，自动识别你有什么：

• Path A（推荐）：把你的 CV PDF、LinkedIn 导出、学历证明、推荐信放进 documents/ 文件夹，自动解析
• Path B：直接粘贴 CV 文本
• Path C：和 AI 进行一次”入职面试”，逐步建立档案

档案拆成 7 个结构化文件：候选人简历（01）、行为特征（02）、写作风格（03）、岗位评估标准（04）、简历模板（05）、Cover Letter 模板（06）、面试准备（07）。

重要：Fork 后必须改成 private 仓库，因为 /setup 会把姓名、联系方式、薪资期望等写入被 git 追踪的文件。

/scrape：搜索职位

同时搜索多个职位平台，去重后按匹配度排序展示。内置平台：Jobindex、Jobnet、Akademikernes Jobbank、Jobdanmark（丹麦市场），以及 LinkedIn（全球）和 freehire.me（多市场）。

其他市场的平台可以用 /add-portal 命令自动生成：给出招聘网站，AI 分析 URL 结构和结果格式，生成并测试新的搜索 skill。

/apply <url>：完整申请流水线

1. 拉取职位描述（无法访问时可粘贴全文）
2. 评估匹配度：五个维度打分，有无 deal-breaker，给出建议
3. 定制简历（LaTeX，lualatex 编译）
4. 写 Cover Letter（LaTeX，xelatex 编译，专用 .cls 文件）
5. 审稿 Agent 批评初稿
6. 修订 → 输出最终版本
7. ATS 可读性检查（需要 poppler，自动降级）

职位描述被视为不可信输入——AI 不会执行其中嵌入的指令，也不会跟随其中的链接。

---

延伸命令

命令	功能
/interview	针对某次面试生成定制准备包：公司研究、可能问题 + STAR 示例映射、模拟面试（roleplay）。不会杜撰经历，缺口给诚实的过渡答案。
/outcome	记录申请结果（面试轮次、offer、拒信、无音讯），归档材料，生成追踪文件。/outcome followup 找出超过 10 天无回音的申请，起草跟进邮件（不发送，至多两次）。
/rank	批量评分所有新抓取的职位，返回排名短名单（并行 Agent 同时处理），过滤截止日期和 dead posting。
/expand	扫描你档案里链接的公开来源（GitHub、portfolio、Kaggle、Google Scholar）发现隐性技能，加入档案并标注来源。
/upskill	分析你的技能与目标岗位的差距，生成优先级热力图和学习计划（含实际学习资源和时间估算）。
/html-report	生成自包含 HTML 仪表板：统计卡、状态/行业/渠道/漏斗图（内联 SVG，无外部依赖），可过滤的申请列表。离线可用。
/notion-sync	单向同步到 Notion 数据库（官方 Notion MCP，OAuth），一行一个职位，只读。
/gmail-sync	读取 Gmail 检测申请状态信号（面试邀请、评估链接、offer、拒信），批量提案供你审核后写入追踪记录。
/add-template	注册自定义简历或 Cover Letter 模板（LaTeX / Typst），测试编译后接入 /apply。

---

技术栈要求

• Claude Code CLI（核心 AI 引擎）
• Python 3.10+
• Bun（职位搜索 CLI 工具）
• LaTeX：TeX Live / MacTeX / TinyTeX（lualatex 编 CV，xelatex 编 Cover Letter）
• 可选：poppler（pdftotext，ATS 可读性检查）

---

为什么值得关注

35k Star 和 12k Fork 的规模不是噶韭菜——这个量级通常意味着真实的使用率。

更有意思的是它的出身：不是某家 AI 公司的 Demo，不是技术博主刷流量的项目，而是一个真的被裁员的人，在真实压力下造的真实工具，用它找到了真实的下一份工作。开源之后，它变成了一个框架，供其他人按自己的市场、自己的简历风格 fork 和改造。

最值得注意的设计选择：整套流程跑在你自己的机器上，数据不离开本地，没有 SaaS 订阅，没有”你的简历数据帮我们训练模型”。你 fork，你填档案，你跑流程，你拥有输出。

---

相关链接

• GitHub：https://github.com/MadsLorentzen/ai-job-search
• 作者 LinkedIn：https://www.linkedin.com/in/mads-lorentzen/
• 视频演示（The Next New Thing）：https://www.youtube.com/watch?v=HoVxjMNFYv4
• Ko-fi（支持作者）：https://ko-fi.com/madslorentzen

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

ai-job-search: 35k Stars — A Geophysicist Built This After Being Laid Off, Got Hired as an AI Engineer

by Mycelium Protocol

---

GitHub: MadsLorentzen/ai-job-search ⭐ 35,452 | Forks 12,170 | Python | MIT
Created: 2026-03-18

---

One Detail First

Mads Lorentzen, a geophysicist, was laid off in late 2025. Instead of updating his LinkedIn, he spent the time building this framework and using it to run his own job search — the same /scrape, /apply, and /interview workflow in this repo, used weekly, on his own career:

• 69 tailored applications
• 20 first interviews
• 1 signed offer
• Started as an AI engineer, June 2026

He told every employer he was using AI assistance in his search. Not once did it count against him — it almost always sparked a genuine technical conversation. He then open-sourced the whole thing. It now has 35,452 stars.

---

An Elegant Recursion

Programmers using their ability to automate things with AI… to find their next job that might be automated by AI. Not as irony — as a pragmatic response to the moment. Job searching is a massively repetitive, highly standardized information processing workflow: read job descriptions, evaluate fit, tailor a CV, write a cover letter, prep for interviews. AI does this more patiently and more consistently than humans, and doesn’t start cutting corners after the 50th application.

---

Core Workflow

/setup        /scrape           /apply <url>
  |               |                  |
  v               v                  v
build profile  search portals    evaluate fit
(from docs /   deduplicate       score + recommend
LinkedIn /     rank by fit
interview)         |                  |
                   v                  v
               pick a match      draft CV + cover letter
               → /apply          (LaTeX, tailored)
                                       |
                                       v
                                 reviewer agent critiques
                                 → revise → final output

/setup — Three paths, auto-detected:

• Path A (recommended): drop your CV PDF, LinkedIn export, diplomas, references into documents/ — automatically parsed
• Path B: paste a CV directly
• Path C: answer an AI “intake interview”

Profile becomes 7 structured files: candidate profile (01), behavioral profile (02), writing style (03), job evaluation framework (04), CV templates (05), cover letter templates (06), interview prep (07).

Important: change the fork to private — /setup writes your name, contact info, and salary expectations into tracked files.

/scrape — Searches multiple portals simultaneously, deduplicates, ranks by fit. Built-in: Jobindex, Jobnet, Akademikernes Jobbank, Jobdanmark (Denmark), LinkedIn (global), freehire.me (multi-market). Other markets: use /add-portal to auto-generate a search skill for any job board.

/apply <url> — Full pipeline:

1. Fetch job description (or paste if blocked)
2. Evaluate fit across five dimensions, flag deal-breakers
3. Draft tailored CV (LaTeX, lualatex)
4. Draft cover letter (LaTeX, xelatex, custom .cls)
5. Reviewer agent critiques the draft
6. Revise → final output
7. ATS parseability check (requires poppler, gracefully degrades)

Job postings are treated as untrusted input — the workflow never follows instructions embedded in them or fetches links from their body.

---

Extended Commands

Command	What it does
/interview	Stage-specific prep: company research, likely questions mapped to your STAR examples, mock interview roleplay. Honest bridge answers for gaps — no invented experience.
/outcome	Record results (interview stages, offers, rejections, silence), archive materials. /outcome followup surfaces applications gone quiet >10 days, drafts a follow-up (max twice, never sends).
/rank	Batch-score all scraped postings in parallel, return a ranked shortlist with per-job strengths and gaps. Filters expired postings and deadline urgency.
/expand	Scan publicly linked sources (GitHub, portfolio, Kaggle, Scholar) for skills not explicit in documents; add to profile with source tags.
/upskill	Analyze skill gaps vs. target roles; produce prioritized gap heatmap and learning plan with actual resources and time estimates.
/html-report	Self-contained HTML dashboard: stat cards, status/sector/channel/funnel charts (inline SVG), filterable table. Fully offline.
/notion-sync	One-way sync to Notion database (official Notion MCP, OAuth). Read-only live view; repo files stay source of truth.
/gmail-sync	Reads Gmail for status signals (interview invites, offers, rejections); proposes as a batch for your approval before anything is written.
/add-template	Register a custom CV or cover letter template (LaTeX, Typst, or any toolchain), with a mandatory test compile.

---

Requirements

• Claude Code CLI (the AI engine)
• Python 3.10+
• Bun (job portal CLI tools)
• LaTeX: TeX Live / MacTeX / TinyTeX (lualatex for CV, xelatex for cover letter)
• Optional: poppler (pdftotext, ATS check)

---

Why It Matters

35k stars and 12k forks at this scale usually indicates genuine adoption, not hype-cycle virality.

More interesting is its origin: not an AI company’s demo, not a content-creator traffic play — a person under real economic pressure who built a real tool, used it to find a real job, and then opened it up for others to fork and adapt to their own market and style.

The most notable design choice: everything runs on your machine. No SaaS subscription, no “your resume data trains our models,” no data leaves your local environment. You fork it, fill in your profile, run the workflow, own the output.

---

Links

• GitHub: https://github.com/MadsLorentzen/ai-job-search
• Author LinkedIn: https://www.linkedin.com/in/mads-lorentzen/
• Video walkthrough (The Next New Thing): https://www.youtube.com/watch?v=HoVxjMNFYv4
• Ko-fi (support the author): https://ko-fi.com/madslorentzen

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.