claude-works: Command for Claude Code

You are onboarding a new candidate onto this job-application system. When you are done, the system must run for THEM: their policy, their search lanes, their resume claims, their identity. Never leave them running on the repo's default (someone else's) policy. Work autonomously; only stop to ask for the inputs listed below and for confirmations that involve their personal facts.

1. Collect the two inputs

Ask the user for:

1. Example jobs they want (3-10 job posting URLs, or titles + companies). ATS URLs (Ashby/Greenhouse/Lever) are best; anything works.
2. Their resume (a file path, pasted text, or a LinkedIn export). Also ask for: name, email, phone, location, LinkedIn/GitHub/website URLs, whether they are authorized to work in the US and need sponsorship, their base compensation floor, any companies they must not apply to (current employer, active interview processes), and any domains they refuse (defense, gambling, whatever they name).

2. Install

pip install -e .        # from this clone; or: pip install claude-works
ruff check . && mypy && pytest   # confirm the install is healthy (all green)
mkdir -p ~/jobsearch-data

3. Derive the personalization (from their examples, not the defaults)

Fetch each example job's JD (fetch_job_description or WebFetch). From the JDs + resume, derive and write:

• ~/jobsearch-data/policy.json (start from examples/policy.sample.json):
  • core_signals: the stack terms that appear in BOTH their resume and the example JDs (their daily work).
  • edge_signals: rare differentiators from their resume (domain experience, unusual skill pairings) that few candidates share.
  • lane_points: the example jobs' title patterns, strongest first.
  • overlevel_terms / level_ok_signals: from their seniority.
  • hard_gap_skills: required-skill terms they lack (ask them to confirm).
  • excluded_companies / excluded_domains / comp_floor: from step 1.
  • seed_boards: the example jobs' org slugs plus 10-20 similar companies (same space, size, stack) that use Ashby/Greenhouse/Lever.
• ~/jobsearch-data/SEARCH_ANGLES.md (format per examples/SEARCH_ANGLES.md): 2-4 lanes distilled from the example jobs; mark the strongest (PRIMARY / default lane).
• ~/jobsearch-data/resumes/_genlib.py: copy examples/resumes/_genlib.py and replace ROLES + every fragment with entries built from THEIR resume. Each fragment must be a claim the user confirms is true; read the list back to them before writing it. Never invent, inflate, or extrapolate a claim. Copy examples/resumes/_render.sh alongside it (it uses Chrome when available).