Job Hunting Is a Funnel Problem, So I Automated It

I'm a lifecycle marketing engineer, which means I spend my working life building automated systems that move people through funnels. In June I ended up on the other side of one: the job application funnel, where I'm the lead being nurtured, scored, and mostly ghosted.

A job search is repetitive, rule-based work with a daily cadence. That's the exact profile of a process you automate. So I built a system that runs my search every morning before I wake up, and this post is about how it works, what it costs me to run (almost nothing), and which parts I deliberately kept manual.

The Repo

The whole system is a git repo called hunter. Inside it: my resume as markdown, a scoring rubric, a watchlist of 98 companies, two Python scripts, and a handful of CSV files that act as the databases. There's no web app and no dashboard. Markdown is the UI and git is the database, because plain files are what an AI agent works with natively, and because I can read a diff and know exactly what happened while I was asleep.

The Problem With Job Boards

Job postings on LinkedIn or aggregator sites are often stale, the role's been filled or pulled and nobody updated the copy. The only reliable source is the company's own hiring page. Checking that directly, every day, across a hundred companies, isn't something a person does by hand, so that's the first thing I automated.

The Part Nobody Told Me About: The Data Is Just Sitting There

Here's what actually solves the problem above. Greenhouse, Lever, and Ashby, the three platforms behind most of these companies' hiring pages, each expose a public API, a direct line to the same live list that builds their careers page. No browser, no clicking through, no waiting for a page to load. Just the current, real answer to "is this job still open."

One of the scripts sweeps my full watchlist of 98 companies against that data every morning, filters for the titles I actually want, and dedupes against everything I've already looked at. The whole pass takes about a minute. A recruiter would call this market coverage. I call it a for loop.

Scoring: The AI Is Not Allowed to Flatter Me

If you paste a job description into an AI chat and ask "am I a good fit?", the answer is yes. It's always yes. That's useless, and worse than useless when you're deciding where to spend application effort.

So the rubric is written down in the repo, and it's mostly penalties. If a role wants five or more years of experience, the score gets capped no matter how good everything else looks. Roles built around platforms I'd need to ramp on lose points. Roles where the salary floor is way above my target range get flagged as a reach, because a posting that's overleveled for me is a rejection with extra steps. AI-forward teams get a small bonus, since heavy AI usage in the day-to-day is the thing I'm actually optimizing for.

Every role that gets scored is logged to a CSV, applied to or not. That log is what makes the daily runs cheap: the system never evaluates the same posting twice, and I never re-read a job description I already decided against.

The Morning Run

Here's where it stops being a script and becomes an employee. Every morning a scheduled cloud agent opens the repo and runs the whole loop from a prompt that lives in the repo itself. It checks the watchlist APIs first. Then it searches the web for roles beyond the watchlist, and this is where the verification rules earn their keep: a role only counts if the agent can find the posting live on the company's own board. Aggregator links don't count. Expired listings don't count. A shocking fraction of what a web search surfaces fails this test.

When the search turns up a company worth watching long-term, the agent suggests it for the watchlist, which is how the list keeps growing without me maintaining it. Then it scores everything new against the rubric, writes a digest sorted into apply first, worth a look, and skipped, and commits the whole thing back to the repo.

I wake up to a commit. I read the digest with my coffee, click through the two or three roles worth a closer look, and decide. The daily grind of a job search, the part that burns people out, takes me about ten minutes.

What Stays Human

Applying is still me. When I decide to go after a role, Claude works like an editor with one hard rule: it is not allowed to invent anything. It reorders and reframes what's already true in my master resume to match what the posting asks for, drafts from my real experience, and everything gets verified against the source file before a PDF goes out. The judgment calls about where to apply, what to emphasize, and what to say to an actual human are mine.

Networking gets the same treatment as everything else: a tracker with the exact message sent, the date, and a follow-up date a week out. Not because outreach should be automated (it shouldn't), but because I will absolutely forget to follow up without a system, and following up is where networking actually works.

What 25 Days of Runs Taught Me

The system has scored 253 roles across 25 consecutive daily runs. Some mornings the digest has three strong matches. Some mornings it has nothing, because that's just the market, and an empty digest is its own kind of signal: I know nothing slipped past me, so I can spend that day on outreach or a side project instead of anxiously re-scrolling the same boards.

The genuinely hard problem wasn't the plumbing, it was getting the scoring honest. Writing the rubric down was the easy half. The real work was the back and forth: testing it against roles I already knew were reaches or easy fits, watching where the scores disagreed with my gut, and tightening the rules until they matched my own judgment. If you build something like this, expect to spend more time here than anywhere else in the system.

And the thing I'd actually pitch to anyone in a search right now: the value isn't speed so much as consistency. The system doesn't have discouraged days. It ran the morning after a rejection email exactly the way it ran the morning after an interview went well. Removing my own morale from the top of the funnel is the single best thing this project did for me.

The Funnel, Closed

There's a version of this post that ends with a grand statement about the future of work. The honest ending is smaller: this is the same job I've been doing for years, take a leaky manual process, wire the systems together, and let humans spend their attention where it counts. This time the funnel was mine.

If you're job hunting and want help setting something like this up, or if you're hiring for lifecycle, marketing ops, or GTM engineering and want this kind of thinking on your team, the form below goes straight to me.