Apply Now (and choose… wisely)

Pharos sweeps a dozen job boards every morning before I’m awake, scores what it finds against my resume, throws out anything that isn’t worth an afternoon, and hands me three or four postings to actually decide on. I’ve been running my own job search on it daily since the start of July.

It exists because the alternative was forty applications a week and two replies, one of them a rejection template that didn’t name the role. The less you hear back, the more broadly you apply. The more broadly you apply, the less any single application is aimed at anything, and that’s exactly what drives the response rate down. A spiral nobody designed, but everybody ends up in it anyway.

Mobile

Desktop

There are tools for this already. Teal and Huntr and Jobscan have been selling versions of it for years at somewhere between thirty and fifty dollars a month⁶. Structurally, it’s the same arrangement LinkedIn Premium is currently under investigation for: people paying a subscription for better odds of not being unemployed. The worse your situation gets, the more the thing is worth to you. That’s a business model with a very unpleasant gradient running through it.

The investigation is real. In July the Texas Attorney General demanded documents⁴ on whether LinkedIn had been selling Premium on the promise of a job marketplace while profiting from listings that weren’t real openings. Nothing has been filed and LinkedIn says its policies require postings to be authentic, so make of it what you will. The underlying number isn’t really in dispute, though: Greenhouse put ghost jobs at 18 to 22% of all online postings², one analysis of LinkedIn specifically landed on 27%³, and the Attorney General’s office puts it somewhere between a fifth and a third. Set against roughly 250 applications per opening¹, that’s a lot of afternoons spent on jobs that were never there.

Hinge has spent years marketing itself as the dating app designed to be deleted⁵, which is a strange promise for a subscription business to make and an entirely coherent thing for a product to aim at. Nobody’s going to build that for job hunting, because job boards are advertising platforms whose business model runs on volume, and a feature that reliably told you which postings were worth your time works directly against the metric they’re measured on.

And while I was building mine, a founder in the States shipped Dreamwork, which ranks live roles against your profile and files applications on your behalf, put together over five months after Indeed laid off his pregnant wife⁷. He landed on nearly the same architecture I had, independently, on the other side of a border. I found that reassuring rather than deflating, because two people staring at the same broken thing and reaching for the same shape of fix is decent evidence that the shape is right.

The one that mattered most was making the resume the rubric. The tool reads what you’ve already written and derives its scoring criteria from that, instead of making you fill in a preferences wizard describing a version of yourself you’d have to invent on the spot. Every spreadsheet tracker I’ve ever abandoned asked me to build the system before I was allowed to use it. This one starts working the moment you hand it something to read.

It also flips the advice everybody gives you: rewrite your resume for every posting you come across. That advice only really makes sense as a coping strategy for spraying applications at volume, and it quietly puts the document on trial in front of each new job description that wanders past. Turning it around so the resume does the judging meant I stopped rewriting myself forty times a week to suit whoever happened to be hiring, which was better for the applications and considerably better for my sanity.

Mobile

Desktop

That screen is where a person is actually in the loop. Everything upstream of it happens without me and is meant to, and everything downstream is me pressing one of two buttons. The whole job of the interface is to get somebody to a decision they’d be willing to defend, in about forty seconds, on reasoning they can argue with rather than a number they have to take on faith. That’s why the domain score sitting at 4 while everything else sits at 5 is the most useful item on it. The machine noticed the same soft spot I would have, and it tells me which part of my background is going to have to do the heavy lifting.

Anything scoring 3/5 or below never reaches the board. That’s a position taken on your behalf before you’ve finished your coffee, and all of it is still sitting there if you want to go and argue with it. The rule underneath that, and underneath everything else, is that nothing it does is allowed to be irreversible. It can surface and score and recommend as much as it likes, but every actual go or no-go stays with you, because you’re the one who remembers getting ghosted by these people last spring.

On the networking side, contacts link through to organizations so you can see the whole picture at any given company, and anything untouched for fourteen days gets flagged as stale. Bad outreach isn’t what kills networking. It’s forgetting to follow up on the good outreach because something else got in the way.

I built all of this with Claude. It’s what happens when a designer with enough technical fluency to specify data structures, parsing logic and product flow gets an engineering partner that can ship overnight. I specified, it built, I used the result the following morning and came back with what was wrong. The daily sweep runs as a scheduled task that spawns its own sub-agents, the dashboards are plain HTML and JSON served by a small Python process, and the integrations run over MCP. None of that is impressive engineering, and it didn’t need to be. It did need someone who understood the problem and could build the structure around it.

The judgment stayed with me. Deciding the resume should be the rubric instead of a preferences form, deciding a 3/5 vanishes rather than sitting on the board making you feel guilty, deciding the actual call gets made at those two buttons and nowhere else. All ordinary design decisions, and all of them took the time design decisions take, because a model will happily build whatever you describe and has no stake in whether your Tuesday morning goes well. The whole thing came together over a few weeks rather than the several months it would once have taken me just to specify it for somebody else to build.

The pipeline keeps narrowing what deserves your attention, and every place where it narrows is a decision about where judgment belongs.

The same logic decides how much effort any one posting has earned. Links only get checked for the ones that already scored 4 or 5, because there’s no sense burning a round-trip confirming something that got thrown out three steps earlier, and the check only fans out across several agents when there are three or more links waiting. Both of those read like engineering optimizations until you notice what they’re actually doing: refusing to spend your morning on things you were never going to apply to.

It started as a single agent working down a list. That stopped being viable the moment there were ten sources to sweep, because a serial run is hostage to whichever board happens to be slowest that morning. So it became a multi-agent architecture, and that turned into its own design problem. The question was no longer what the AI shows the user, but the stranger one of how several AI actors should coordinate before anything gets shown at all.

The split follows the judgment. Collecting postings is commodity work with nothing to weigh, so it fans out across agents that never need to know about each other, while scoring and deduplication and presentation all carry judgment and come back through a single agent holding the whole picture. The human checkpoint sits at the end rather than somewhere in the middle, so you get asked to decide once against a short list instead of repeatedly against a long one.

I stopped applying to everything and started evaluating fit first. That sounds obvious written down, and it took a piece of software to actually make it happen. The rubric gave me something I hadn’t expected: permission to skip a posting without guilt. When the reasoning is sitting right there in front of you, it’s much harder to agonise about whether you’re being too picky. It also gets there first more often than not: on a typical morning three out of five new postings are already scored and waiting by the time LinkedIn decides to send its daily digest. And networking stopped being a permanent shapeless guilt and became a pipeline with states in it. I’d have called that a trivial change if the staleness flag hadn’t gone on to alter my behaviour more than every advice article I’ve ever read put together.

The hardest part of designing for AI is the seam between the algorithm and the person standing in front of it. A score of 4/5 is an opinion, it might be wrong, and you have to hold both of those at once without concluding the whole thing is useless. A narrow gap to aim at: show too little of the reasoning and you’ve built a black box; show too much and you’ve buried the decision in the middle of it.

Trust gets earned in unglamorous ways: reasoning you can actually read, overrides that take one click, and some grace about the times it gets things wrong, which it will.

I built Pharos because nothing else worked and I needed something that did. It turned out to be the most instructive design problem I’ve had my hands on in years, because the material itself — non-deterministic outputs applied to decisions that actually matter to somebody — demands a kind of thinking that no amount of making the button bigger will ever get you to.

Which is, in the end, the name. It throws light on the water: the scores, the reasoning, the postings that were never real jobs. Then it stops there, because a lighthouse doesn’t steer the ship.