job-hunting

Job Hunting

Use this skill when a user wants help searching for jobs from a resume, homepage, GitHub profile, or other candidate material.

Workflow

1. Intake candidate material.

   • Accept a local resume file, raw resume text, homepage URL, or GitHub profile URL.
   • Do not expect the user to provide target roles, highlighted projects, locations, or search preferences upfront.
   • If the user provides only candidate material, infer the candidate's strongest options and ask the next questions yourself.
   • Use jobhunter.profile.load_input to normalize the input into text.
   • Use jobhunter.profile.build_user_profile to build a structured UserProfile.

2. Inspect and refine the profile.

   • Check roles, skills, locations, seniority, links, and keywords.
   • Use profile.evidence["portfolio_highlights"] to inspect the strongest resume/homepage/GitHub points.
   • Use adaptive_preference_questions(profile) to identify missing user preferences.
   • Ask focused intake questions before search, in this order when relevant:
     • Full-time, internship, or both. Ask this early — it is one of the first questions. Never infer it silently: a resume saying "PhD student" or "research intern" does NOT mean the candidate wants an internship. Record the answer in job_types.
     • Which portfolio highlights should be emphasized most.
     • Which role tracks should be prioritized from the inferred options.
     • Preferred locations and remote/hybrid/onsite preference.
     • Preferred industries, company stages, salary floor, and visa needs.
     • How many jobs to return.
     • Output format. Default to TSV when the user does not choose a format.
   • Present the questions as concrete choices derived from the profile, not as generic open-ended homework for the user.
   • Record answers with update_preferences and save_preferences, then rebuild or reuse the profile with those preferences.
   • If the profile is clearly missing core facts, ask one concise clarification before running a broad search.
   • If the user only asked for intake, stop after the questions and wait for answers.
   • If the user wants autonomous progress or explicitly asks to search immediately, proceed with inferred preferences and state the assumptions. Default to full-time when the user has not answered — internship postings are ranked lower (set job_types="full-time" to make this explicit), unless the candidate clearly signals they want internships.

3. Select sources.

   • Prefer free structured sources first: RemotivePlatform and RemoteOKPlatform for remote roles, HackerNewsWhoIsHiringPlatform for HN Who's Hiring.
   • Use GreenhousePlatform and LeverPlatform for direct employer boards (AI labs and strong engineering brands by default; tune the company lists with JOBHUNTER_GREENHOUSE_BOARDS / JOBHUNTER_LEVER_COMPANIES to fit the candidate's targets).
   • Use CompanyBoardsPlatform (--sources companies) to search employers' own career sites. By default it sweeps 100+ verified company boards (OpenAI, Anthropic, xAI, SpaceX, Databricks, Snowflake, Netflix, Uber, Stripe, Pinterest, Thinking Machines Lab, and more). To focus on specific employers, set JOBHUNTER_COMPANIES (or .jobhunter/companies.json) to a comma list of friendly names, e.g. openai,anthropic,databricks; add a not-yet-listed board with greenhouse:<slug> / lever:<slug> / ashby:<slug>. Companies without a public API (Meta, TikTok/ByteDance, Salesforce, PayPal) are skipped with a warning; use --sources linkedin for LinkedIn.
   • Use LinkedInJobSpyPlatform when the user wants LinkedIn and python-jobspy is installed under Python 3.10+. LinkedIn silently returns zero rows when rate-limited; do not treat an empty LinkedIn result as "no jobs".
   • Use A16ZPlatform for a16z portfolio jobs and YCJobsPlatform for YC startup jobs.
   • Check PLATFORM_REGISTRY for implemented and planned sources.
   • Treat Glassdoor and Handshake as planned/restricted until a compliant adapter exists.
   • Use OfflineDemoPlatform only for tests, demos, or no-network situations.
   • Add paid, authenticated, or restricted platforms only behind the JobPlatform interface and only through compliant access paths.

4. Run agentic search.

   • Use QueryPlanner to generate role/skill/location queries from the profile.
   • Use JobSearchAgent to search each selected platform, normalize postings, deduplicate, rank, and produce explainable matches.
   • Always pass --trajectory-log logs/trajectory-<date>.jsonl for live runs so the search is auditable (queries, per-source fetch/keep counts, errors, final ranking).
   • Refine by adding/removing query terms when the result set is too broad, too sparse, or mismatched.
   • Retry loop: if fewer results than requested came back, or fewer than 3 sources contributed, rerun with broader queries (--agentic-search), higher --per-query-limit, more --max-queries, or relaxed target roles before reporting a sparse list.

5. Verify quality before presenting.

   • Run "$CLAUDE_PLUGIN_ROOT/scripts/evaluate" <run.json> --check-urls --report <eval.md> on the JSON output.
   • The evaluator scores four metrics: platform diversity, freshness (posting age + live URL checks), profile match, and duplication.
   • If a metric reports NEEDS ATTENTION, iterate (broaden sources, tighten preferences, drop stale results) rather than presenting a weak list.

6. Present results.

   • Lead with ranked jobs, company, source, location, salary if available, URL, and why it matched.
   • Every result carries apply_url (direct application link) and outreach_message (a profile-tailored greeting); surface both so the user can apply and reach out immediately, and offer to refine the message tone per company.
   • Offer an HTML view (--html > jobs.html) when the user wants something nicer than a table: it produces a self-contained, browsable page with apply buttons, matched-term chips, a filter box, and one-click-copy outreach messages. Save it to a file and tell the user the path (or open it).
   • Always report: which sources contributed and which returned nothing or errored, freshness caveats (e.g. YC postings carry no dates), and any assumptions made on the user's behalf.
   • Call out source limitations, especially for unstructured HN comments or platforms without public APIs.
   • Offer concrete follow-ups: verify shortlisted jobs are still open before applying, export the shortlist (TSV/JSON), draft tailored resume bullets or outreach notes for top matches, add sources, or record preferences for the next run.

Optional Storage

Persistence is disabled by default; runs are stateless. To record jobs and search runs across sessions, set JOBHUNTER_DATABASE_URL (SQLite or Postgres) in the environment before running — see the README's storage section. Never require storage for a basic search.

Running the Engine

Always invoke the engine through the bundled launcher, never a bare python:

"$CLAUDE_PLUGIN_ROOT/scripts/jobhunter" <args>

CLAUDE_PLUGIN_ROOT is set by Claude Code to this skill's install directory. When it is unset (e.g. running from a clone), use the repo-relative path scripts/jobhunter. The launcher is zero-config: it finds a Python 3.10+ runtime automatically and runs the dependency-free engine with no venv and no pip. The user never sets up Python. The only optional dependency (python-jobspy, for the LinkedIn source) is provisioned on demand through uv; if uv is missing, every source except LinkedIn still works. The quality evaluator has its own matching launcher at scripts/evaluate.

Useful Commands

Offline deterministic demo:

"$CLAUDE_PLUGIN_ROOT/scripts/jobhunter" --input examples/resume.md --offline --limit 5

Live Remotive demo:

"$CLAUDE_PLUGIN_ROOT/scripts/jobhunter" --input examples/resume.md --sources remotive --limit 5

Startup boards and direct company pages:

"$CLAUDE_PLUGIN_ROOT/scripts/jobhunter" --input examples/resume.md --sources a16z,yc,companies --limit 10

Preference questions and recorded answers:

"$CLAUDE_PLUGIN_ROOT/scripts/jobhunter" --input examples/resume.md --offline --show-questions
"$CLAUDE_PLUGIN_ROOT/scripts/jobhunter" --input examples/resume.md --preferences-file .jobhunter/preferences.json --set-preference job_types="full-time" --set-preference target_roles="research scientist,machine learning engineer" --set-preference preferred_locations="Remote US,New York" --set-preference remote=true --sources a16z,yc,companies

JSON output and quality evaluation:

"$CLAUDE_PLUGIN_ROOT/scripts/jobhunter" --input examples/resume.md --sources remotive,companies --json > run.json
"$CLAUDE_PLUGIN_ROOT/scripts/evaluate" run.json --check-urls --report eval.md

LinkedIn (python-jobspy is auto-provisioned via uv; no manual install):

"$CLAUDE_PLUGIN_ROOT/scripts/jobhunter" --input examples/resume.md --sources linkedin --limit 10

Skill Invocation Examples

Minimal GitHub portfolio intake:

$job-hunting https://github.com/CurryTang

Minimal resume intake:

$job-hunting /path/to/resume.pdf

Expected behavior:

• Analyze the material first.
• Show inferred roles, skills, locations, and portfolio highlights.
• Ask early whether the candidate wants full-time, internship, or both; default to full-time and rank internships lower when unanswered.
• Ask the user which inferred highlights and role tracks to prioritize.
• Ask location, remote, industry, stage, salary, and visa questions only when missing.
• Ask how many jobs to return and which output format to use; use TSV by default.
• Do not start search until the user answers, unless the user explicitly asks for autonomous search.