I built a free interview-prep generator, then discovered it was prepping people for the wrong call

I run a small career-tools site as a side project. One of the free tools generates likely interview questions from a pasted job description. People used it, so I assumed it worked.

Then I actually audited my own output. I fed it a JD that literally said "Competitive salary. Remote (EU)." and it returned 15 questions — none about pay, none about remote logistics, and it led with an "Answer using the STAR method" panel. I checked the whole question bank, not just that draw: of the 53 questions the tool can produce, zero mention salary, compensation, notice period, start date, work authorization, or "why are you leaving."

Here's the thing I got wrong: the entire interview-prep genre — mine included — is built for the hiring-manager round. But the first call is almost always a recruiter screen, which is a completely different filter: comp range, logistics, and "can this person hold a normal conversation." STAR-drilling actively hurts you there — recruiters are listening for whether you sound rehearsed.

So the tool was technically working and strategically useless for the exact moment most candidates hit first. Classic case of building for the demo, not the workflow.

I'm reworking it to detect screen-stage vs. manager-stage from the JD and branch the questions. If you're job hunting the tool is free and needs no signup — but honestly the more useful takeaway for anyone building tools is: audit your own output against a real input before you trust your own product. I'd used it for months without noticing.

Tool (free, no signup): https://charliemorrison.dev/interview-prep/?utm_source=indiehackers&utm_medium=post&utm_campaign=career&utm_content=interview-prep

Curious if others have caught their own tool solving the wrong version of the problem — how'd you notice?