Screen thousands of job postings and write fact-checked tailored resumes

Chavez Lopez, an enterprise architect, built his own job-search agent on a Mac Studio using a Claude Pro plan he already paid for. A scheduler runs it in Docker with SQLite for storage. It pulls postings from employer career systems such as Greenhouse, Lever, Ashby, Workday and SmartRecruiters, and uses Playwright for sites without a feed. Nine coded filters drop wrong titles, stale posts and unsupported locations before any model sees them. Claude Haiku then screens the rest against a 103-keyword skills index, each keyword backed by evidence in his 30 KB career history. For roles that pass, it rechecks the posting is still open, and Claude Sonnet writes a tailored resume and verifies every claim against the career file. He reviews and applies himself; nothing is submitted automatically.

In one test cycle it took 7,710 raw postings, matched 1,342 on title and produced 12 worth his time. Earlier versions went wrong in ways he had to fix: an off-the-shelf agent recommended jobs that had already closed, feeding the whole career file made almost every role look like a strong match, and drafts used inflated corporate wording. He added the availability recheck, the skills index, an 80% match threshold and stricter writing rules. The extra recurring cost is $0. He gives no figures yet on interviews or offers.

Most of the savings come from doing the filtering in plain code before calling a model, which also keeps him inside the plan's usage limits. Pointing it at employer career systems rather than job-board copies is what stops dead listings. The fact-check step can still miss things, which is why he keeps the final review manual.