What the “harness” actually is
Under every matched role is a quiet checklist the model runs before it spits out a score. That checklist is the harness — the set of questions we ask the AI on your behalf each time a new job shows up: Does the title fit? Does the work look like what this person has done well? Are the must-haves really there, or only implied?
Most AI products hide this layer. We exposed it on purpose. A grocery store manager, a seventh-grade teacher, and a nonprofit program lead care about very different signals, and the defaults cannot read minds. What they can do is be adjustable in language anyone can use.
Three plain-English dials, no code anywhere
You do not write prompts, you do not touch JSON, and you never need to learn what a model temperature is. You just tell the system what matters and what does not. There are three levers, and the most important one starts with five signals to teach your job search scorer:
- Rule learning — Atlas learns from your feedback. Mark jobs interested or dismissed and Atlas synthesizes scoring rules that boost roles like the ones you liked and de-prioritize ones like the ones you dismissed.
- Draft Rules — short written guardrails like “ignore roles that require a relocation to a high-cost city” or “prefer schools with 500+ students.” The AI applies them on the next run and shows you which matches would have changed.
- Feedback deltas — every thumbs up or thumbs down nudges the weights. Do it enough times and the top of your list starts to look like the jobs you would have picked anyway, only found faster.
Why curiosity is enough
The reason this works for non-technical users is that the tuning loop is transparent and reversible. Before any change goes live, you see a preview: which of yesterday’s roles would have moved up, which would have fallen, and why. If it feels wrong, you undo it. Nothing you do here breaks anything.
That is a different promise from most AI tools, which ask you to trust the output. Here the deal is simpler — the model does the heavy lifting, and you keep steering.
What tuning changes in a real week
After a handful of rounds, two things shift. First, the scores become more honest for your specific situation — an 88 means 88 to you, not to a generic “job seeker.” That is what AI job match scoring is supposed to become when the model has enough of your feedback. Second, the reasoning sentences next to each role start to read back your own language, because you taught the harness that language.
The outcome is not magic. It is fewer tabs open, fewer afternoons re-sorting the same listings, and a shortlist you actually trust well enough to act on in the morning.