Stop applying to ghost jobs. A free, open-source AI agent skill family that searches LinkedIn, Indeed, and state job boards, scores each posting 1-5 for legitimacy and CV-fit, and outputs a tailored, ATS-ready resume + cover letter per role. Built for every career: nurses, welders, teachers, engineers. Runs cross-OS (Windows / macOS / Linux) in Claude Code, OpenAI Codex, OpenClaw, Hermes Agent, and any agentskills.io-compatible agent.
v6.0.0 is now a skill family. What used to be one large skill is an orchestrator (
job-hunter) that routes to five focused member skills — so "just tighten my resume" runs the resume skill alone, while "run my whole job search" fans out across all of them. The orchestrator owns the shared.job-hunter/workspace and the anti-fabrication safety gate, and decides which members each request needs.
Skill Role What it does job-hunter orchestrator Routes your request to the right members; owns the workspace + the anti-fabrication invariant career-profile profile-intake Captures your North-Star profile (5 questions) + parses your resume job-search search Finds, dedupes, and scores postings 1-5 across 4 search tiers resume-tailor resume-tailor Tightens (Mode A) or tailors (Mode B) your resume, with the verify-no-fabrication gate application-tracker tracker Tracks applications + drafts stale-application follow-ups outcome-learning learning-loop Learns from your outcomes; proposes opt-in lessons Install them together with the family installer (Install below).
People are looking for jobs and changing career paths all the time. The job search is full of noise. Ghost listings that were never real, irrelevant matches, generic advice, hours wasted tailoring resumes for roles that already had a hire lined up. It eats time you don't have.
I also believe agent skills are the next big breakthrough in agentic AI. Small, focused, composable capabilities you bolt onto whatever agent you're already using. Combining the two was the obvious move.
This is what job-hunting looks like when an AI agent actually does the boring parts. Free, open-source, runs locally, never uploads your resume.
Once installed (see Install below), just ask your agent:
Find me senior backend engineer jobs in Seattle, $180k+, that aren't ghost listings.
Here's my resume, what nurse practitioner jobs near Austin would I be a strong fit for?
Tailor my resume for this posting: [URL]
The skill auto-activates, walks you through 5 quick profile questions on first use (saved to .job-hunter-profile.md in your workspace, never sent anywhere), then runs the full pipeline:
- Understand the user, North-Star profile (5 questions) + resume parsing
- Search, LinkedIn, Indeed, Glassdoor, ZipRecruiter, Dice, Wellfound, USAJobs, 50-state workforce commissions, and company career pages
- Score, every posting gets a 1-5 across 5 sub-scores (CV match, comp vs target, cultural signals, posting legitimacy, red flags)
- Tailor, ATS-optimized DOCX resume + cover letter per role you want to pursue
- Track, interactive HTML tracker with sort/filter, sub-score breakdown, and weighted global score
Pick your harness. All four install paths are first-class.
| Harness | Skills directory | Install guide |
|---|---|---|
| Claude Code | ~/.claude/skills/ |
install/claude-code.md |
| OpenAI Codex | ~/.codex/skills/ |
install/codex.md |
| OpenClaw | ~/.openclaw/skills/ |
install/openclaw.md |
| Hermes Agent | ~/.hermes/skills/ |
install/hermes.md |
If you just want it working in 10 seconds:
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/wexxwuther/job-hunter/main/install/install.sh | bash# Windows (PowerShell)
iwr https://raw.githubusercontent.com/wexxwuther/job-hunter/main/install/install.ps1 -UseBasicParsing | iexPrefer not to clone? Grab the latest release zip from the Releases page and extract it into your agent's skills directory. Done.
Every job posting gets scored across five sub-scores (1-5 each) with documented weights:
| Sub-score | Weight | What it measures |
|---|---|---|
cv_match |
0.35 (heaviest) | How well your resume aligns with the posting's stated requirements |
comp_vs_target |
0.25 | Listed comp vs your target band from your North-Star profile |
cultural_signals |
0.20 | Mission, team size, hiring-page tone, glassdoor signal |
posting_legitimacy |
0.20 | Ghost-job rubric (re-posts, vague specs, missing recruiter) |
red_flags_penalty |
multiplier | Pays-in-equity, asks-for-SSN-pre-offer, fee-to-apply, torpedoes the score |
A single severe red flag can take an otherwise-perfect posting from 5.0 to 1.0. Read the full rubrics in references/posting-legitimacy-rubric.md and references/match-quality-rubric.md.
job-hunter/
├── SKILL.md # The skill itself (agentskills.io spec)
├── scripts/ # 15 production Python scripts (no network deps)
│ ├── score_posting.py # Deterministic 1-5 scoring
│ ├── parse_resume.py # DOCX/PDF/text resume parser
│ ├── extract_ats_keywords.py # Keyword-gap analysis
│ ├── build_search_queries.py # Multi-board query builder
│ ├── expand_role_synonyms.py # Role-title synonyms (SDE = SWE = Engineer)
│ ├── normalize_salary.py # $120k/yr = $10k/mo = $60/hr
│ ├── dedupe_postings.py # Cross-board dedup
│ ├── generate_tracker_html.py # Interactive HTML tracker
│ ├── init_profile.py # North-Star profile management
│ ├── init_workspace.py # Sets up the .job-hunter/ learning loop
│ ├── harvest_outcomes.py # Reads outcomes, finds patterns
│ ├── propose_lessons.py # Suggests lessons for you to confirm
│ ├── draft_followup.py # Check-in + thank-you email drafts
│ ├── export_workspace.py # Bundle your state to a zip
│ └── import_workspace.py # Restore from a zip on a new machine
├── references/ # 9 reference docs the skill consults
├── assets/templates/ # Template files + tracker CSS
├── tests/ # 177 unit tests
└── evals/ # Trigger + outcome eval suites
No API keys required. No telemetry. No phoning home. Your .job-hunter-profile.md stays in your workspace.
Job-hunter keeps a small, per-user memory of decisions, outcomes, and confirmed lessons, all in plain-markdown files in your workspace, never sent anywhere.
File (lives in <your-workspace>/.job-hunter/) |
What it captures |
|---|---|
DECISIONS.md |
Per-session choices and why (skipped Acme because location, chose light tailoring for Beta) |
LESSONS.md |
Patterns about your preferences, you confirm each one before it's added |
OUTCOMES.md |
What actually happened (accepted, rejected after onsite, no response after 21 days) |
REJECTED_IDEAS.md |
Hard constraints, "no defense contractors", "no commission-only", agent never re-asks |
The cycle: as you apply and outcomes land, scripts/harvest_outcomes.py looks for patterns. When ≥5 closed-loop outcomes exist, scripts/propose_lessons.py translates the signals into suggested lessons. The agent surfaces each suggestion with evidence and asks "want me to remember this?", only confirmed lessons get added to LESSONS.md.
Six guardrails keep it honest:
- Cold-start guard: no lessons proposed until ≥5 closed-loop outcomes (no pattern detection from thin data)
- Opt-in only: the agent never auto-writes lessons; every entry is user-confirmed
- Deterministic translation: same input → same suggestion; reasons are anchored in evidence, not paraphrased
- Bounded influence: lessons adjust how sub-scores are graded for you; the scoring weights in
score_posting.pystay constant for all users - Plain markdown: every file is readable, editable, deletable by you, no black-box weights
- Local-only: no telemetry, no phone-home, no upload
Delete .job-hunter/ to start over. The skill re-initializes it on next run. See references/learning-loop-guide.md for the full design and trust model.
cd job-hunter
pip install pytest # only dependency
pytest tests/
# 177 passedCI runs the full suite on every push across Python 3.10/3.11/3.12 on Ubuntu, macOS, and Windows.
I don't have a resume yet, can I still use this?
Yes. If no resume is in your workspace, the skill will offer two paths: (1) point to one you have (file upload, path, or paste), or (2) walk you through a focused interview to draft a baseline resume.md from scratch, work history, accomplishments with numbers, skills, education, projects, and any non-traditional experience. Once the baseline is solid, the normal job-search and tailoring flow kicks in. The skill won't run the job search with no resume at all, because the scoring rubric and tailoring step both need one.
My resume is thin or out of date. The skill will flag that before tailoring rather than producing 10 weak tailored versions of a weak base resume. It'll walk you through a short strengthen-the-baseline pass (asking for numbers on your accomplishments, surfacing skills you haven't listed, filling in dates) and then proceed.
Will this submit applications for me? No. It produces the tailored resume + cover letter + application tracker. You submit. That's a deliberate boundary, auto-submission is a different trust model and out of scope for this skill.
Does it send my resume or profile anywhere?
No. Your resume, your .job-hunter-profile.md, and every tailored output live in your workspace. The skill has no telemetry, no phone-home, no required API keys, and no required network calls beyond what your agent already does to search the web. See SECURITY.md for the full trust model.
Why isn't there an interview-prep feature? Deliberately out of scope. Interview prep is a different workflow with different inputs and a different success criterion. If you want it, run a separate interview-prep skill alongside this one.
Does the skill learn from my interactions?
Yes, but entirely locally and opt-in. After 5+ closed-loop application outcomes, the skill surfaces patterns it noticed (e.g., "4 of 5 rejections cite comp, want me to remember this?") and asks you to confirm before adding any lesson. Confirmed lessons live in <your-workspace>/.job-hunter/LESSONS.md and adjust how sub-scores are graded for you. The scoring weights themselves stay constant. Nothing is uploaded or used for training. See the learning loop section above and references/learning-loop-guide.md.
Can I reset what the skill has learned about me?
Yes, delete <your-workspace>/.job-hunter/. The next run will re-initialize empty templates. You can also edit any of the four files directly (LESSONS.md, DECISIONS.md, OUTCOMES.md, REJECTED_IDEAS.md) since they're plain markdown.
Will the skill help me follow up on applications?
Yes. scripts/draft_followup.py has two templates: check_in (7-10 days after applying with no response) and thank_you (24-48 hours after an interview). The skill scans your tracker.json for applications stale at status=applied for 7+ days and offers to draft polite check-in emails. It also drafts thank-you notes after you report an interview. The skill never sends the email, that's by design, and a load-bearing safety test enforces it. You copy-paste. See references/followup-templates.md for the patterns and cited sources.
Can I move my job-hunter state to a different machine?
Yes. scripts/export_workspace.py bundles your profile + .job-hunter/ learning loop + tracker.json (and optionally your tailored DOCX outputs) into a single zip. scripts/import_workspace.py restores them on a new machine. The export refuses to write to cloud-sync directories (Dropbox, OneDrive, iCloud) by default; pass --allow-cloud if you intentionally want that. Roundtrip preserves all files byte-for-byte. The import path-sanitizes the archive to reject any traversal payloads.
Does this work for non-tech careers?
Yes, this is the explicit design goal. references/niche-boards-by-industry.md includes verified entries for healthcare (Vivian Health, Nurse.com, Health eCareers, AlliedTravelCareers, state nursing board portals), trades (iHireConstruction, AllTrucking, IBEW/UA local hall search patterns, SkillBridge for military-to-trades), legal (LawCrossing, BCG Attorney Search, state bar career pages), education (HigherEdJobs, SchoolSpring, K12JobSpot), government (USAJobs, GovernmentJobs.com, NEOGOV-hosted municipal portals, state civil-service exams), and more. The agent picks the right tier-2 boards based on your role.
I asked it to tighten my resume and it added things I didn't say. What went wrong?
That's a real bug we fixed in v5.2.0. Earlier versions of Phase 3 had truth-preservation listed as a Key Principle near the bottom of the SKILL.md, while the actual phase instructions said things like "push for numbers and outcomes." LLMs follow active instructions over passive guardrails, so the skill added quantification and content the user hadn't given it. v5.2.0 splits Phase 3 into two explicit modes: Mode A (Tighten/Copyedit) is a zero-fabrication contract that may not add numbers, change job titles, or pull content from external websites, and Mode B (Tailor for a posting) runs a mandatory verification gate (scripts/verify_no_fabrication.py) that surfaces every new claim before any output file is written. If you ever see the skill add something you didn't say, that's still a bug; please open an issue.
Issues, feature requests, and PRs welcome. See CONTRIBUTING.md for the workflow. If you found a security issue, see SECURITY.md for private disclosure.
If you use job-hunter to land a job, open a Discussion and tell us, that's the highest-signal feedback this project gets.
MIT. Use it commercially, fork it, rebrand it, ship it inside your product. No attribution required (but appreciated).
- Competitive design review against
santifer/career-ops(MIT),job-hunterv4 folded in four of its ideas (1-5 scoring, legitimacy/match split, North-Star profile, multi-block tracker) and rejected five others where our design constraints differ. - Built on the open agentskills.io spec so it runs across Claude Code, Codex, OpenClaw, and any future compatible agent.