--- name: job-seeker description: >- Manages LaTeX resume variants for job applications: tailor bullets to a JD, compile PDFs, convert between Markdown/DOCX/PDF/LaTeX, recommend variants, run browser-assisted apply (dry-run/rehearse only unless user confirms submit), update the application tracker, and enforce honesty boundaries. Use when the user mentions 申职器, job-seeker, resume, JD, LaTeX compile, convert resume format, apply to a job posting, or tracker updates. --- # job-seeker — 申职器 Workflow skill for multi-variant LaTeX resumes plus format conversion and browser-assisted application (human-in-the-loop). Enforces honesty boundaries so bullets never overstate scope, metrics, or titles. ## Editor install paths | Editor | Path | Notes | |--------|------|-------| | **Codex** | `.codex/skills/job-seeker/` + `AGENTS.md` | Committed project Skill + repo-level entry | | **Cursor** | `.cursor/skills/job-seeker/` | Committed — EN + `SKILL.zh-CN.md` | | WorkBuddy | `.workbuddy/skills/job-seeker/` | Add `agent_created: true` to frontmatter | | Portable copy | `docs/skill/job-seeker/` | Mirror for export / generic editors | Codex setup guide: `docs/skill/CODEX.md`. Cursor setup guide: `docs/skill/CURSOR.md`. Chinese skill: `SKILL.zh-CN.md` (same directory). Windows build: `docs/BUILD_WINDOWS.md` · compile with `.\build_resumes.ps1`. ## When to use Trigger when the user: - Pastes a JD and wants a tailored resume or variant recommendation - Asks to compile / rebuild resume PDFs (`./build_resumes.sh`) - Wants format conversion (`python -m convert ...`) - Wants apply dry-run / rehearse (`python -m apply --dry-run ...`) - Edits `docs/experience_bank.md` or resume `.tex` bullets - Updates the application tracker after submitting - Asks whether a bullet violates honesty boundaries - Asks about 申职器 / job-seeker Do **not** trigger for generic LaTeX help unrelated to this repo's resume workflow. ## 1. Repo map (relative paths) ``` ./ # job-seeker / 申职器 ├── resume_template/ # Public sample template + bundled CJK fonts (always present) │ └── sample-resume-en_US-zh_CN.tex ├── LaTeX_Resume_CN/ # YOUR private CN variants (create locally; gitignored) ├── LaTeX_Resume_EN/ # YOUR private EN variants (create locally; gitignored) ├── convert/ # python -m convert — MD/DOCX/PDF/LaTeX any-to-any ├── apply/ # python -m apply — JD→variant, prefill, confirm gate ├── docs/ │ ├── experience_bank.md # Private bullet library (gitignored — copy from .example) │ ├── experience_bank.example.md │ ├── jd_mapping/ # Private JD notes (gitignored) │ ├── BUILD_MAC.md │ └── skill/job-seeker/ # This skill (portable mirror) ├── outputs/ # Compiled PDFs + tracker (gitignored except README) ├── history/ # Apply audit logs (gitignored) ├── build_resumes.sh # macOS/Linux — compiles LaTeX_Resume_* if present ├── build_resumes.ps1 # Windows equivalent ├── scripts/sync_skills.ps1 # Windows — sync skill mirrors ├── scripts/sync_skills.sh # macOS/Linux — sync skill mirrors ├── .codex/skills/job-seeker/ # Codex project skill ├── AGENTS.md # Agent entry (Codex / Cursor / compatible editors) ``` **First-time setup:** copy `resume_template/` files into `LaTeX_Resume_CN/` or `LaTeX_Resume_EN/`, personalize, then run `./build_resumes.sh`. See root README Quick start — do not commit personal `.tex` to the public repo. ## 2. Standard workflow ### A. New JD → tailored resume 1. Read JD (pasted text, URL, or `docs/jd_mapping/.md`) 2. Read `docs/experience_bank.md` — **do not invent bullets** 3. Recommend variant: `python -m apply --jd-file --dry-run` or use `references/variant_playbook.md` 4. **Plan A (default):** copy general variant `.tex` → change only `\centerline` headline → save as `resume-zh_jd_.tex` 5. Build: `./build_resumes.sh` (Windows: `.\build_resumes.ps1`) or compile single file in `resume_template/` 6. Verify page counts in script output — JD variants target **1 page** 7. Update tracker via `apply` audit or manually append `outputs/job_application_tracker.csv` ### B. Edit bullets or skills 1. If multi-resume impact, edit `docs/experience_bank.md` first 2. Apply to each affected `LaTeX_Resume_*/*.tex` 3. `./build_resumes.sh` — check `resume_job_en` stays **1 page** ### C. Format conversion ```bash python -m convert --input path/to/resume.tex --output outputs/resume.pdf python -m convert --list-routes ``` - **Source of truth for layout:** `.tex → .pdf` via `xelatex` - LaTeX ↔ DOCX is **lossy** — warn user; macros/icons do not survive - PDF → Markdown extracts text only (no OCR for scanned PDFs) Details: `convert/README.md`. ### D. Apply (rehearse / prefill) ```bash export JOB_SEEKER_NAME="Your Name" export JOB_SEEKER_EMAIL="you@example.com" python -m apply --jd-file apply/samples/sample_jd.html --dry-run python -m apply --url https://example.com/jobs/123 --rehearse ``` **Iron rules:** - `submit()` always blocks on human `confirm()` — **never auto-submit** - Boss直聘 / LinkedIn: `tos_blocks_automation=True` → open page + show data only - Credentials via env vars only — never commit to repo Details: `apply/README.md`. ### E. After submission - Tracker row in `outputs/job_application_tracker.csv` (schema: `docs/tracker_schema.md`) - Optional: `history/apply_.log` via `apply` audit - Optional: cover note in `outputs/cover_notes/.txt` ## 3. Variant selection | Variant | Typical source `.tex` | Output PDF | When | |---------|----------------------|------------|------| | Sample / demo | `resume_template/sample-resume-en_US-zh_CN.tex` | same dir | First compile, CI | | CN general | `LaTeX_Resume_CN/resume-zh_job.tex` | `resume_job_cn.pdf` | Domestic dev (1 pg) | | CN backend/ops | `LaTeX_Resume_CN/resume-zh_backend_ops.tex` | `resume_backend_ops_cn.pdf` | Ops / SRE | | CN AI/RAG | `LaTeX_Resume_CN/resume-zh_ai_eval.tex` | `resume_ai_eval_cn.pdf` | AI / RAG roles | | CN JD | `LaTeX_Resume_CN/resume-zh_jd_.tex` | `resume_jd__cn.pdf` | Headline-only JD | | EN general | `LaTeX_Resume_EN/resume_job_en.tex` | `resume_job_en.pdf` | English roles (1 pg) | | EN backend/ops | `LaTeX_Resume_EN/resume_backend_ops.tex` | `resume_backend_ops.pdf` | EN ops | | EN AI/RAG | `LaTeX_Resume_EN/resume_ai_eval.tex` | `resume_ai_eval.pdf` | EN AI eval | | Web3 | `LaTeX_Resume_EN/resume_web3.tex` | `resume_web3.pdf` | Anonymous crypto | Full openers/DM templates: `references/variant_playbook.md`. ## 4. Honesty boundaries ### §4.1 Content authenticity - No invented metrics — numbers must trace to real work - No inflated titles (e.g. "Technical Lead" ≠ "Staff Engineer") - "Maintained/extended existing platform" ≠ "built from scratch" - ML/retrieval: cite real parameters (TopK, thresholds, chunk strategy) - POC (few workflow nodes) ≠ "autonomous agent" or production system ### §4.2 Skills-line discipline - Skills = languages, dev toolchain, general AI productivity tools — short list - Do **not** put project stacks (RAG, Kafka, etc.) in Skills — put in project bullets - AI assistants (Cursor, Codex, Trae) are **tools**, not programming languages ### §4.3 Open-source project framing - Describe side projects by what they actually do — neutral, accurate domain - GitHub URL on its **own line** (avoid PDF line-wrap overflow) - Do not reframe a research toolkit as live trading / order routing / signals product ### §4.4 Length - JD variants + EN general: **1 page** (`enumitem` + tight bullets) - Full CN with research section: multi-page OK ### §4.5 Drafter–reviewer pass (model-agnostic) When tailoring a variant to a JD, do **two passes**, not one. This is pure prompt craft — it works under any editor (Cursor / Codex / WorkBuddy) and requires no paid subscription or specific model. 1. **Drafter.** Produce the tailored `.tex` variant from the chosen template + experience bank, cutting/reweighting bullets for this JD. Keep it honest (§4.1). 2. **Reviewer (separate turn / fresh context).** Re-read ONLY the JD + the drafted variant. Critique, do not flatter: - Weak verbs, clichés, filler ("utilized", "leveraged", "various"). - Bullets that read stronger than the underlying work (§4.1 inflation). - JD keywords present in name only (keyword-stuffing without substance). - Orphans, overflow, >1 page where §4.4 says 1 page. - Missing quantification where real numbers exist. 3. **Revise.** Apply only the critiques that keep the bullet truthful. If a critique would inflate, reject it and note why. The reviewer pass is advisory — it never auto-edits. It runs the same honesty rules as the drafter; its job is to catch what a single pass misses. You can also run the heuristic honesty linter on any variant before sending: ```bash python -m apply --honesty-check LaTeX_Resume_EN/sample-resume-en_US-zh_CN.tex ``` It flags vague superlatives, absolute claims, and scope-inflation phrasing (honesty boundary §6.1). Advisory only — confirm against checklist.md §E. ### 4.6 Skill-gap analysis (the "what am I missing?" view) Beyond tailoring, help the candidate see where they fall short for a role. Run the skill-gap analysis on the JD (and optionally their current resume): ```bash python -m apply --upskill [--resume LaTeX_Resume_EN/.tex] ``` It reports the JD keywords the resume does **not** contain, framed as skills to develop — the complement of the ATS keyword *coverage* check. Use it to: - Tell the candidate honestly which bullets to add (if the experience is real). - Flag gaps that would require fabrication — and advise against inventing them (honesty boundary §6.1). The tool never suggests lying; it suggests either acquiring the skill or not applying where the gap is fundamental. ## 5. Build & verify ```bash chmod +x ./build_resumes.sh && ./build_resumes.sh ``` Windows (PowerShell): ```powershell .\build_resumes.ps1 ``` See `docs/BUILD_WINDOWS.md` or `docs/BUILD_MAC.md`. Critical page-count checks when `LaTeX_Resume_*` exists: - `resume_job_en.pdf` → **1 page** - `resume_jd_*_cn.pdf` → **1 page** If `LaTeX_Resume_CN/` missing, script fails — user must create variants first or compile `resume_template/` directly with `xelatex`. Missing `xelatex` → exit 2, see `docs/BUILD_MAC.md`. ## 6. References | File | When to load | |------|--------------| | `references/master_prompt_extract.md` | Full vision + extended rules | | `references/variant_playbook.md` | Variant pick + opener templates | | `checklist.md` | Before submit / before PR | | `docs/skill/CODEX.md` | Codex-specific setup, prompts, MCP boundary | | `docs/skill/CURSOR.md` | Cursor-specific tips | | `convert/README.md` | Conversion routes & lossy warnings | | `apply/README.md` | Apply adapters & confirm gate | ## 7. Placeholders (public repo / examples) Use `YOUR_NAME`, `your-email@example.com`, `your-org/job-seeker`, `YOUR_OPEN_SOURCE_PROJECT`, `COMPANY_X` — never real maintainer PII in examples. ## 8. Agent behavior (Codex / Cursor) 1. **Read before write:** `docs/experience_bank.md` + target `.tex` 2. **Minimal diff:** do not refactor `convert/` or `apply/` core unless asked 3. **No silent commit/push** — user must request explicitly 4. **Relative paths only** — no `D:\project\...` or absolute home paths in docs 5. **Verify builds** after `.tex` edits when `xelatex` is available 6. **Apply safety:** default to `--dry-run` / `--rehearse`; live submit only with explicit user `y` ## 9. Out of scope - Auto-submitting applications without user confirmation - Committing personal data (real tracker rows, experience bank, private `.tex`) - OCR for scanned PDFs (not implemented) - Modifying resume bullets without reading `docs/experience_bank.md`