Imported from BaoAL31/CVTailor (AGENTS.md). Install upstream with npx skills add BaoAL31/CVTailor. Copyright stays with the author.

Agent Instructions

Short workflows for this repo. Detail lives in docs/.

Workflow

Discover (find jobs)

# Defaults: LinkedIn only, last 24h (data/search_defaults.yaml)
python -m src.discover --limit 15

# Include other boards/ATS alongside LinkedIn
python -m src.discover --sources linkedin,prosple,gradconnection,greenhouse,lever,ashby --limit 15

# Repeat discover+scrape every 2h (starts Chrome :9223 if LinkedIn is included)
python -u scripts/discover_loop.py

1. Keywords/location from CLI or data/search_defaults.yaml
2. Each keyword × each --sources → auth → search
3. Skip senior/lead/staff titles; skip rows already in jobs/<date>/discovered_jobs.csv or data/seen_jobs.db
4. Scrape new listings into jobs/<YYYY-MM-DD>/<company>-<title>/job-description.md
5. Merge new rows into jobs/<date>/discovered_jobs.csv
6. Tailor → apply for folders the user picks

Do not pass --easy-apply unless asked (default off). Full behaviour: docs/discover.md.

Tailor + apply (known job)

When the user provides a JD (or URL):

1. If URL only → python -m src.scrape_job <url> (docs/scraping.md)
2. Read master CV CV_master.docx and the JD
3. Ensure folder jobs/<company>-<title>/ with job-description.md
4. Generate in that folder:
   • CV_Tailored.docx — medium-intensity tailor (rules below)
   • Cover_Letter.docx — template below
5. Apply — python -m src.apply jobs/<folder>/ --run for LinkedIn Easy Apply and SEEK Quick Apply (docs/apply.md). Other ATS: docs/apply/README.md without browser MCP.
6. Write handoff.md summarising the session

Chain: discover → scrape → tailor → apply

Apply — CDP drivers (no MCP)

Do not use Playwright/browser MCP to apply. Do not open new tabs or launch Chromium. Debug Chrome must already be running on :9222 (python scripts/hold_chrome.py or .\\launch-linkedin-chrome.ps1). Attach only.

python -m src.apply jobs/<date>/<folder>/ --run
python -m src.apply jobs/<date>/<folder>/ --run --no-submit   # stop at Review

LinkedIn Easy Apply and SEEK Quick Apply are implemented. Other processes: follow docs/apply/README.md using the same Chrome (no MCP), or extend src/apply_*.py.

For ATSes without a driver (Workday etc.): do NOT connect() per step. Every connect() leaks a CDP session and Chrome stops answering after ~5–6. Hold ONE attachment in a server and drive it command-by-command:

python -m src.browser.wd_server            # attach once; serves 127.0.0.1:9223
python -m src.browser.wd_client --op ping  # unlimited commands, no re-attach

Same page object stays alive → SPA state is stable and the connection budget never degrades. Details + full op list: docs/apply/workday.md.

Shared CDP/inspect/answer-bank rules: docs/apply/README.md and docs/apply.md.

Pre-flight (every job)

python -m src.generate CV_master.docx jobs/<folder>/job-description.md --output-dir jobs/<folder>/
python scripts/list_apply_candidates.py   # optional

Confirm the listing is still open. Skip if "No longer accepting applications" — --mark-status expired.

--run re-reads the page after each action. Agents must not drive apply via MCP snapshots.

Answer bank

python -m src.apply_answers lookup "<exact question text>"   # NOT_FOUND if missing
python -m src.apply_answers save "<exact question text>" "<answer>"

For ANY personal/identity/qualification field — name, title, address, phone, degree, major, school, dates, skills, years of experience, visa/eligibility, salary, DOB, gender — ALWAYS run apply_answers lookup FIRST and use the bank's exact value. Never type these from scratch, never abbreviate, never substitute. A bank match that is worded differently still wins (e.g. the bank says degree = "Bachelor of Computing Science (Honours)" — use exactly that, NOT "Bachelor"). If a required field has no bank match, STOP and ask the user — never guess. Park unknowns in jobs/<folder>/pending_questions.yaml if batching. After first use, save every new Q&A.

Default screening answers (only when consistent with the CV — never invent skills/years): see docs/apply/linkedin-easy-apply.md and the answer bank.

Confirm + record

1. Confirm the submission: success message (e.g. "Application submitted" / "Your application has been sent to " / "Thank you for applying").
2. Record in the master ledger jobs/applications.csv + seen_jobs.db. For rows with source/external_id in frontmatter, python -m src.apply jobs/<date>/<folder>/ --mark-applied (or --mark-status skipped|done|rejected|expired --notes "<why>" for non-applied). Seek / GradConnection rows often lack the id — append to jobs/applications.csv directly.
3. Write handoff.md in the job folder (template below).

handoff.md template

# <Company> — <Title>

- **Status:** Applied | Skipped (reason) | Blocked (reason)
- **Date:** YYYY-MM-DD
- **URL:** https://...
- **CV:** CV_<Name>.docx (+ cover letter if uploaded)
- **Answers:** brief note on non-default screening answers

Apply checklist (copy for agents)

• CV + cover letter generated in job folder
• Listing still accepting applications
• Form completed through submit (--run for LinkedIn/SEEK; no MCP)
• Unknown questions asked / parked — not invented
• Success message confirmed
• Recorded in ledger (applied or skipped/done) + handoff.md written
• New answers saved to data/apply_answers.yaml

Personal Context

Use these for application screening answers (seeded in data/apply_answers.yaml; extend via apply_answers save).

Tailoring (LLM-based)

• Master CV is the only source of truth — LLM never invents skills/experience
• LLM (NVIDIA NIM, model minimaxai/minimax-m3; falls back to OpenRouter) rewrites bullet points, selects 2–3 most relevant projects, and reorders skills
• Allowed: remove irrelevant entries, reword bullets to echo JD keywords, reorder by relevance
• Keep dates, companies, and role titles accurate; header stays identical to master
• Falls back to original CV if LLM call fails (no silent corruption)
• Full project-blurb context: docs/portfolio.md

Cover Letter Template

• Starts with "Dear Hiring Manager," (no header block)
• Body: 3–4 paragraphs connecting CV to JD
• Close: Best regards, [full name]
• Signature: phone | email | GitHub