job-seeker

job-seeker — 申职器

Languages: English · 简体中文

Tailor a clean LaTeX resume to any job description in minutes — then convert it between Markdown / DOCX / PDF / LaTeX, and optionally prefill job applications behind a hard human-in-the-loop safety gate. Ships with bundled CJK fonts (compiles out of the box) and an Agent Skill that teaches your AI editor the whole workflow.

MIT · cross-platform (macOS · Linux · Windows) · LaTeX + pandoc + Playwright

How job-seeker is different. Most AI job-search tools bundle a paid agent subscription, scrape job boards, and auto-pilot submissions. job-seeker goes the other way: offline, no subscription, scrapes nothing, and never auto-submits — and it is built script-agnostic: every writing system is a first-class citizen — Han (CJK), Latin, Cyrillic, Arabic, Hebrew, Devanagari, Greek, and more. Bring a resume in any format and any language; get a clean LaTeX / PDF / DOCX with pixel-perfect CJK typography (the hardest case we solve best — if xelatex + CJK fonts work, the rest is font + bidi work, not a rewrite), and stay in control. Works with whatever AI editor you already use (Cursor / Codex / WorkBuddy).

✨ Why job-seeker?

Sending the same resume to every job gets you filtered out. Hand-tailoring one per JD is slow, keeping .tex / DOCX / Markdown in sync is painful, and submitting applications is error-prone. job-seeker turns that into one coherent flow:

• Bring any format, get any format. Start from .tex, .md, .docx, or .pdf — convert freely between all four without manual sync.
• JD-aware variant picking. Paste a JD, get a recommended resume variant.
• Apply without losing control. A browser assistant prefills forms but never auto-submits — every submission waits for your explicit y.
• Honest resumes by design. A built-in Skill enforces anti-inflation rules (no invented metrics, no inflated titles).

🚀 Features

• Clean LaTeX template with bundled Adobe CJK fonts — no system font install needed.
• Cross-platform build scripts (build_resumes.sh / .ps1) with page-count reporting.
• Any-to-any format conversion (convert/): LaTeX ↔ Markdown ↔ DOCX ↔ PDF — upload a resume in any of the four formats and convert to any other, via an adapter pattern with automatic BFS routing (PDF text is ingested with PyMuPDF).
• JD → variant recommender (apply/): rule-based keyword matching picks the right resume for the role.
• Browser-assisted application with site adapters (Boss直聘 / LinkedIn / generic ATS) and a human-in-the-loop confirm gate.
• Application tracker (CSV) + history audit log — every action recorded.
• Agent Skill (Codex / Cursor / WorkBuddy / editor-agnostic) that encodes the workflow and honesty boundaries.
• CI compiles the sample resume on every push/PR.

🧭 How it works

flowchart TD
    subgraph ANY["1 · Bring any format"]
        direction LR
        tex[".tex"]:::fmt
        md[".md"]:::fmt
        docx[".docx"]:::fmt
        pdf[".pdf"]:::fmt
    end
    conv["convert/ — any-to-any conversion<br/>10 direct edges · BFS routing"]:::core
    ANY --> conv
    bank["experience_bank.md<br/>(private, gitignored)"] --> vars["LaTeX_Resume_CN / EN<br/>(your variants)"]
    conv --> vars
    jd["Job description<br/>URL · paste · file"] --> rec["recommend_variant<br/>(rule-based)"]:::core
    vars --> rend["render_variant → PDF"]:::core
    rec --> rend
    rend --> gate{"confirm()<br/>human-in-the-loop"}:::gate
    gate -- "y" --> app["apply/<br/>prefill form — never auto-submit"]:::core
    gate -- "N" --> cancel["cancel"]
    app --> track["tracker.csv + history/"]:::core

    classDef fmt fill:#eef,stroke:#88a,color:#225;
    classDef core fill:#e9f7ec,stroke:#5b9,color:#163;
    classDef gate fill:#fdeef0,stroke:#e88,color:#721;

Four formats interconvert freely; PDF enters as text-only (PyMuPDF). The apply path is gated — no submission ever happens without your explicit y.

📦 Installation

Prerequisites

Setup

git clone https://github.com/Cavaradossi/job-seeker.git
cd job-seeker
python3 -m venv .venv && source .venv/bin/activate   # Windows: .\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

⚡ Quick start

1. Compile the sample resume

cd resume_template
xelatex -interaction=nonstopmode sample-resume-en_US-zh_CN.tex
# → sample-resume-en_US-zh_CN.pdf (CJK fonts included, no extra setup)

2. Convert between formats

# The input format is inferred from the extension — start from ANY of .tex / .md / .docx / .pdf
python -m convert --input resume_template/sample-resume-en_US-zh_CN.tex --output outputs/sample.pdf
python -m convert --input resume_template/sample-resume-en_US-zh_CN.tex --output outputs/sample.docx
python -m convert --input my_resume.pdf   --output outputs/my_resume.md    # PDF → Markdown (text extracted with PyMuPDF)
python -m convert --input my_resume.docx  --output outputs/my_resume.tex   # DOCX → LaTeX
python -m convert --list-routes           # see every supported conversion

3. Tailor to a JD & rehearse the apply flow

python -m apply --jd-file apply/samples/sample_jd.html --dry-run    # read JD → recommend variant → render PDF
python -m apply --url https://example.com/jobs/123 --dry-run        # fetch a live JD from the web
python -m apply --jd-file apply/samples/sample_jd.html --rehearse   # feel the confirmation gate pause

💡 Usage examples

Make a JD-specific variant. Copy the general template, change only the headline, compile:

cp resume_template/sample-resume-en_US-zh_CN.tex LaTeX_Resume_CN/resume-zh_jd_acme.tex
# edit \centerline{...} headline to match the JD, then:
./build_resumes.sh

Convert a Markdown draft to PDF (uses pandoc's default template — not a resume layout, handy for drafts):

python -m convert --input notes.md --output notes.pdf

Apply to a public ATS posting (generic adapter; live prefill needs Playwright):

export JOB_SEEKER_NAME="Your Name"
export JOB_SEEKER_EMAIL="you@example.com"
python -m apply --url https://boards.greenhouse.io/acme/jobs/123 --variant resume_job_en
# → browser opens, fields prefilled, then it PAUSES for your [y/N] before submitting

⚙️ Configuration

• Apply profile — set via env vars (never stored in the repo):
  • JOB_SEEKER_NAME, JOB_SEEKER_EMAIL, JOB_SEEKER_PHONE
  • JOB_SEEKER_CONFIRM=yes|no — auto-answer the confirm gate (for scripting/tests only)
• Resume variants — drop your .tex files into LaTeX_Resume_CN/ or LaTeX_Resume_EN/; the renderer searches those plus resume_template/.
• Tool discovery — convert/tools.py finds pandoc / xelatex / latexmk on PATH first, then falls back to common install locations (MacTeX, ~/texlive/*/bin/*, Homebrew, ~/.local/bin), so the CLI works even in non-interactive shells.
• ToS-aware adapters — Boss直聘 and LinkedIn set tos_blocks_automation = True, so apply degrades to "open page + show prefill data" instead of automating them.

📁 Project structure

job-seeker/
├── resume_template/        Clean LaTeX template + bundled CJK fonts + fontawesome
├── convert/                Phase 2: format conversion pipeline (MD/DOCX/PDF/LaTeX)
│   ├── adapters/           One adapter per conversion (pandoc / xelatex)
│   ├── pipeline.py         BFS routing + chained adapters
│   └── cli.py              python -m convert ...
├── apply/                  Phase 3: browser-assisted application
│   ├── adapters/           Site adapters (generic / boss_zhipin / linkedin)
│   ├── mapping/            JD→variant recommender + variant renderer
│   ├── confirm.py          Human-in-the-loop gate (submit() always blocks here)
│   ├── audit.py            Tracker CSV + history log
│   └── cli.py              python -m apply ...
├── docs/                    Build guide, open-source plan, tracker schema, example bank
├── docs/skill/job-seeker/  Portable Agent Skill source (sync target)
├── .codex/skills/job-seeker/  Codex project skill
├── .cursor/skills/job-seeker/  Cursor project skill (open repo → ready)
├── .workbuddy/skills/job-seeker/  WorkBuddy project skill
├── build_resumes.sh/.ps1   Cross-platform build scripts
└── requirements.txt        Python deps for convert/ + apply/

❓ FAQ

Do I need to install system fonts? No. The template bundles Adobe CJK fonts in resume_template/fonts/, so it compiles identically on any machine.

Do I need pandoc? Only for DOCX / Markdown conversion. tex → pdf works with just xelatex.

Will it auto-submit my job applications? Never. submit() always calls confirm() and blocks. Nothing is submitted without an explicit y. For Boss直聘 / LinkedIn (whose ToS prohibit automation), it doesn't even try — it opens the page and shows you the data to paste manually.

Is my personal data safe in the public repo? Yes by design. outputs/ (compiled PDFs + tracker CSV), history/, docs/experience_bank.md, and your real .tex variants are all gitignored. The public repo contains only the clean template + code, using YOUR_NAME / your-email@example.com placeholders.

Which editors does the Skill support? Codex — project skill is under .codex/skills/job-seeker/; start from AGENTS.md and docs/skill/CODEX.md. Cursor — open the repo; skill is under .cursor/skills/job-seeker/ (English + Chinese). See docs/skill/CURSOR.md. WorkBuddy loads from .workbuddy/skills/. Sync after edits: .\scripts\sync_skills.ps1 (Windows) or ./scripts/sync_skills.sh.

Can I start from a PDF or Word doc instead of LaTeX? Yes. Upload any of .tex / .md / .docx / .pdf and convert to any other — the pipeline routes automatically. PDF is text-only on the way in (extracted with PyMuPDF): a scanned/image PDF has no text layer and will be rejected, and layout/styling is not recovered. Treat PDF-in as "recover the content, then re-typeset with the LaTeX template."

The conversion to DOCX looks lossy — why? LaTeX ↔ DOCX is fundamentally lossy: custom resume macros (\name, \centerline, \datedsubsection, icons) don't survive. The DOCX is an editable text mirror, never a layout-accurate render. The .tex → .pdf path is the source of truth. (The CLI prints a WARNING on every lossy step.)

🤝 Contributing

Contributions are welcome — and there's plenty to do! A few good first areas:

• New site adapter for an ATS you use (Workday / Greenhouse / Lever variants welcome) — see apply/adapters/workday_generic.py as a template.
• New conversion adapter (e.g. HTML → PDF, or OCR for scanned PDFs) — implement the Adapter ABC in convert/adapters/.
• More JD → variant rules in apply/mapping/jd_to_variant.py.
• Better field-detection heuristics for the prefill step.

Before opening a PR:

1. Read the open-source plan for direction and the privacy strategy.
2. Respect the honesty boundaries in docs/skill/job-seeker/SKILL.md §6 — no resume inflation, no invented metrics.
3. Add a test under convert/tests/ or apply/tests/ and run pytest (33 tests should stay green).
4. Use placeholders (YOUR_NAME, your-email@example.com) in any example content — never real personal data.

⭐ Star History

📄 License

MIT — see LICENSE.

The resume_template/ directory is derived from billryan/resume (also MIT). All bundled fonts retain their original licenses.