Job Hunter

Job Hunter Skill

Automates your job search: scrapes LinkedIn, Indeed, and 100+ job boards daily, scores every posting against your resume, and saves a ranked tracker to Obsidian. No LinkedIn account needed.

Quick Start

First-Time Setup

1. Set up virtual environment:

   bash ~/.claude/skills/job-hunter/scripts/setup_venv.sh
   

2. Configure preferences:

   ~/.venv/job-hunter/bin/python3 ~/.claude/skills/job-hunter/scripts/setup_config.py
   

3. Set your resume path (PDF, DOCX, or TXT)

4. Set Obsidian vault path

Usage

Run the job hunt:

/job-hunt

Or with specific options:

/job-hunt --keywords "software engineer" --location "San Francisco"

How It Works

Two complementary providers run on a smart schedule:

On days both run, results are merged and deduplicated. On off-days, JobSpy runs alone.

Workflow

When invoked, the skill executes these steps:

1. Load Configuration

• Read config from ~/.config/job-hunter/config.json
• Validate paths and optional API keys

2. Read Resume

• Load resume from configured path
• Extract key skills for matching

3. Search Jobs — JobSpy (daily) + JSearch (every 2 days)

JobSpy hits LinkedIn's undocumented public guest API — no login, no API key, returns full job descriptions (~60 results/run).

JSearch queries Google Jobs via RapidAPI, aggregating postings from LinkedIn, Indeed, Dice, Remotive, and 100+ company career pages (~100 results/run).

The scheduler uses a state file (~/.job-hunter-jsearch-last-run) to decide whether JSearch runs. If both run, their outputs are merged and deduplicated:

JobSpy  ~60 jobs ─┐
                  ├─ deduplicate → ~130–160 unique jobs
JSearch ~100 jobs─┘

4. Score against resume

Each job is scored 0–100 based on skill keyword matching:

• +5 per skill match (Python, SQL, machine learning, etc.)
• Bonus: role alignment (+15), healthcare domain (+10), EHR tools (+8), remote (+5)
• Skills are extracted automatically from your resume (PDF/DOCX/TXT)

4b. AI Re-rank (interactive runs only)

Keyword scoring is literal — it can't distinguish "we use Python for Excel macros" from "ML platform team". When the skill runs interactively (not via cron), Claude re-ranks the top jobs by actually reading them:

1. Prep — build a compact packet of the top ~30 keyword-scored jobs + resume:

   ~/.venv/job-hunter/bin/python3 ~/.claude/skills/job-hunter/scripts/rerank.py prep \
     -i /tmp/scored.json -o /tmp/rerank_packet.md --top 30
   

2. Judge — Claude reads /tmp/rerank_packet.md and scores every job 0–100 for holistic fit (seniority, domain, day-to-day work vs. resume — not keyword counts). Claude writes the result to /tmp/rerank_scores.json as:

   {"<job id>": {"ai_score": 88, "ai_reason": "Strong healthcare + Python overlap"}}
   

   Every job id from the packet must appear. Keep reasons under 80 chars.
3. Apply — merge scores back and re-sort:

   ~/.venv/job-hunter/bin/python3 ~/.claude/skills/job-hunter/scripts/rerank.py apply \
     -i /tmp/scored.json -r /tmp/rerank_scores.json
   

Each job gains final_score (AI score when present, keyword score otherwise) and the tracker sorts/filters by it, adding a Why column. Cron runs skip this step automatically — write_tracker.py falls back to keyword scores.

5. Save Job Tracker to Obsidian

Writes Job Tracker - YYYY-MM-DD.md with all jobs scoring > 70, sorted by score descending. Idempotent — skips if today's file already exists.

| # | Score | Title | Company | Salary | Remote | Source | Link |
|---|-------|-------|---------|--------|--------|--------|------|
| 1 | 100   | Data Engineer II | Dana-Farber Cancer Institute | - | Yes | Indeed | [Apply](url) |
| 2 | 100   | Senior Health Data Informaticist | Veeva Systems | - | Yes | Linkedin | [Apply](url) |

6. Wait for User Selection — STOP HERE

After saving the tracker:

"Saved X jobs to Job Tracker - YYYY-MM-DD.md in Obsidian. Review the file and tell me which numbers you'd like cover letters for (e.g. '1, 3')."

Never auto-generate cover letters. Always wait for user to select.

7. Generate Cover Letters

~/.venv/job-hunter/bin/python3 ~/.claude/skills/job-hunter/scripts/generate_cover_letters.py \
  --jobs /path/to/scored_jobs.json \
  --indices "1,3"

Saves to Cover Letters/YYYY-MM-DD/Company - Title.md in your Obsidian vault.

Configuration Reference

Config file: ~/.config/job-hunter/config.json

{
  "jsearch_api_key": "your-rapidapi-key",
  "search_provider": "jobspy",
  "resume_path": "/path/to/resume.pdf",
  "obsidian_vault": "~/Documents/Obsidian Vault/job-hunter",
  "user_name": "Your Name",
  "min_score": 70,
  "search": {
    "keywords": ["data analyst", "healthcare data analyst", "research analyst"],
    "location": "Boston, MA",
    "remote": true,
    "time_range": "week",
    "max_results_per_query": 20,
    "job_domains": ["linkedin.com", "indeed.com"],
    "linkedin_fetch_description": true,
    "request_delay": 3,
    "proxies": [],
    "user_agent": "",
    "country_indeed": "USA"
  }
}

Key Fields

Scripts

Obsidian Structure

job-hunter/
├── Job Tracker - 2026-02-28.md   # Today's tracker
├── Job Tracker - 2026-02-27.md   # Yesterday's tracker
├── Cover Letters/
│   ├── 2026-02-28/
│   │   ├── Dana-Farber - Data Engineer II.md
│   │   └── Veeva Systems - Senior Health Data Informaticist.md
│   └── 2026-02-27/

Scheduling

Set up a daily cron to run the full pipeline automatically:

# Run job hunt daily at 10am
0 10 * * * bash ~/.claude/skills/job-hunter/scripts/daily_job_hunt.sh

The script decides internally whether to also run JSearch (every 2 days). To force a JSearch run:

rm ~/.job-hunter-jsearch-last-run

API Usage & Cost

Troubleshooting

No jobs found

• Check ~/.job-hunter.log — the run now prints per-site totals (Per-site totals: linkedin=0, indeed=25) so you can see which board failed
• Broaden search keywords in config
• Each site is scraped independently, so one board failing no longer wipes out the others' results

LinkedIn returns 0 results / 429 errors

• LinkedIn aggressively rate-limits (~100 results per IP, then blocks). In config:
  1. Set search.linkedin_fetch_description: false (biggest win — cuts requests by ~95%)
  2. Lower search.max_results_per_query
  3. Increase search.request_delay (e.g. 10)
  4. Add search.proxies — per JobSpy maintainers, proxies are the only durable fix
• Indeed has no rate limiting and full descriptions — results still flow while LinkedIn cools down (blocks typically clear within hours)

JSearch API errors

• Verify key at rapidapi.com
• Check monthly usage (200 free requests)
• Delete ~/.job-hunter-jsearch-last-run to retry

Glassdoor 400 errors in logs

• Known upstream issue in JobSpy — Glassdoor is now excluded from default sites; remove glassdoor.com from job_domains if you added it

Virtual environment issues

• Re-run bash ~/.claude/skills/job-hunter/scripts/setup_venv.sh
• Requires Python 3.10+