The half of Compass I did not need to build

Compass is the system I run my job search out of, and it takes as its input a list of companies I put together by hand. That hand-built list nagged at me, so the obvious next question was whether it could build itself. I gave it two or three working hours, spread across a couple of weeks, to find out, on a separate thing I called CareerOS, which was supposed to interview me about what I actually want out of a company, go find companies that matched, and hand back a ranked list of twenty-five without me curating anything.

It got built and it ran, with a questionnaire, a resume reader, a catalog of invented companies to test against, and a ranked list that put a plain-English reason under every company alongside an honest note about what it still did not know. I want to be careful about how much credit that deserves, because standing up a working app and a pile of research to justify it is close to free now, and the volume of either proves nothing at all. Producing it was never going to be the hard part, and treating it as the accomplishment is how you end up shipping something confident and useless.

The only test that counted was pointing it at my own resume, and it told me I was an executive. The bug is funnier than it deserves to be, because it was scanning for "cto" as a job title and the word "director" has c-t-o sitting right in the middle of it. I fixed that, and a better problem was waiting behind it, since even reading correctly the thing reported the most senior title I had ever held rather than my current one. Nobody had told it those were two different questions, and nobody was me. That turned out to be the shape of the whole project, where every answer it handed back was only as good as the question I had thought to ask, and I was quietly supplying all of the judgment while the software supplied all of the confidence.

So I killed it, and I am not open-sourcing it. The target list was never my bottleneck, because I already know which companies I care about and have known for a while. Finding out whether a human being sits inside one of them is close to table stakes too, since LinkedIn will tell you that much for free. What Compass actually earns its keep on is keeping a deliberately narrow list of the places I most want to work in front of me instead of buried in a spreadsheet, curating the outreach that goes against it, and pulling the job boards, the last time I spoke to someone, and the tasks that follow from both into one place so that running a search does not have to live in my head. That is the part worth tuning, and the part worth writing down here is not that I built the other thing, it is that two or three hours and one real input were enough to find out I should not have. So the next stretch of work is Compass itself, tuned against a search that is actually running, which is the only place the remaining questions are going to come from.