Agentic Job Search — AI skill for Claude Code

agentic-job-search

An agentic job-application system I built with [Claude Code](https://claude.com/claude-code) while running my own Summer 2027 internship search. It watches 46 companies' job boards directly, ranks every posting by evidence-based fit, verifies I'm actually qualified before a job reaches my list, scores my resume against a public hiring rubric, and gates every application behind five automated checks.

**This system never auto-submits.** It exists to make me apply to *fewer* jobs, more accurately: postings are verified (open, qualified-for, honestly matched) before I see them, and a five-gate check blocks any application whose resume claims can't be evidenced. Every application is individually reviewed and submitted by hand.

Built AI-first: I directed the design and Claude Code wrote and iterated the implementation across agentic sessions. Every classifier decision in here traces to a real failure we hit and fixed — the commit history and code comments keep those receipts.

What it does

**`watch/`** — the job watcher, run hourly by cron. Deterministic: no model in the loop, so polling often is free.

• Polls Greenhouse, Ashby, Lever, Workable, and Workday APIs directly (aggregators lag those same endpoints by hours to days), plus community aggregator lists as a slower second net, deduplicated against the direct feeds by canonical URL.
• Scores each posting 0–100 — role fit, employer signals, term confirmation, freshness — divided by application effort read from the ATS (a Greenhouse form costs minutes; a Workday account-creation slog does not).
• Reads the full job description, not just the title: postings that require a PhD/Masters (without a bachelor's alternative) never reach the apply tiers. Titles lie by omission; requirements sections don't.
• Verifies links are alive before listing them, with care: some boards ship "Page not found" as template text on every live page, so a page showing the job's own title is alw