A parent with zero GitHub experience built a job-search agent that runs daily

Project Description: - Achievements: I used Claude to architect this agent through conversation — the search logic, the exclusion taxonomy, and a privacy safeguard (contact details used for the session, never stored) — then deployed it through Claude Code as a genuinely autonomous Routine. It has run unattended every morning at 8:00 a.m. since it went live, with zero ongoing maintenance from me. That's the real achievement: a working autonomous system, built end-to-end through conversation with Claude by someone with zero prior GitHub or Routines experience — not a one-off script, and not something a no-code platform assembled for me. A sample run (August 3, 2026: 12 postings, correctly filtered and deduped) is attached as evidence of what it does, not as the achievement itself. - Tools: Claude (chat — architecture and prompt design), Claude Code (Routines — scheduled cloud deployment), GitHub (a private repo holding the state-tracking file), Gmail connector (delivery). - Methodology: My younger son graduated from Embry-Riddle's Prescott campus in May 2026, into what's arguably one of the worst aviation job markets in recent memory. He'd already done two internships at Las Vegas-area airports and was manually grinding through applications — close to a hundred of them, no interviews — targeting Las Vegas, Phoenix, and Denver, where he has friends. Watching him get worn down by the search itself, not just the rejections, I opened a conversation with Claude to design a fix. We worked out the mechanism together: a private GitHub repo holding a single state file (seen_postings.md) so the agent never reports the same posting twice; a detailed search prompt covering LinkedIn, Indeed, Glassdoor, and airline/airport career pages — prioritizing Allegiant Air, Frontier Airlines, and other carriers based in the three markets — against the right taxonomy (operations & dispatch, ground ops, safety/compliance, management trainee tracks, analytics, UAS/drone, cargo/logistics), while excluding anything requiring a pilot certificate, ATC certification, or a Part 107 remote pilot cert; and a Gmail connector to deliver results, or an explicit "nothing new" note, to both my son and me. I had zero GitHub experience going in — Claude walked me through creating the account, the repo, and the file from scratch. There was a real wrong turn along the way: the desktop app defaulted me into a local session that wanted Git installed, which had nothing to do with what I actually needed. Fixing it meant understanding the difference between a local session (my machine) and a cloud session (Anthropic's servers) — Routines only need the cloud version. Once that was sorted, I built the routine through Claude Code's Routines interface, pointed it at the repo, connected Gmail, set it to run daily at 8:00 a.m., and tested it immediately with a manual run rather than waiting a full day to find out if it worked. - Learnings: The hardest part of agentic filtering isn't finding postings — it's teaching the agent what to correctly ignore, and what to remember it's already shown you. Without the state-tracking, a "daily" search is just the same list restated on a loop. I also learned a real agent build isn't one prompt — it's a short back-and-forth to get the architecture right, including deciding upfront not to have either of our contact details stored permanently, before the routine ever runs on its own. - Relevant URLs: Internal/personal — not public-facing.