Building an AI resume-tailoring tool that doesn't trust itself

This started small — tailoring my own resume for one application, going back and forth until it felt right. But I kept saying the same things over and over: don't make up numbers, don't use a title I can't defend, don't list something I only saw once in a course. So I stopped repeating myself. Turned it into an actual process built around the thing that made the first attempt work in the first place — it asks something, I answer, and it checks that answer against everything else before moving on.

None of the rules were just guesses, either. Most resume advice floating around online is folklore that's been repeated so many times it sounds like fact — "the ATS bot will auto-reject you" is the big one everyone's heard. Before I wrote that part of the tool I actually looked into how recruiters at big companies use their hiring software day to day, instead of trusting the myth. Turns out nothing auto-rejects you. A human looks, always. The actual risk is being invisible in a search, not getting bounced by a robot. That's more or less the standard for every rule in here — check what's true first, write the rule after.

Then I actually used it on real applications, and it fell apart in a way I didn't see coming. Not the writing — the process itself. It would skip a step and just not say anything. Answer its own confirmation question instead of waiting for mine. Ask something I'd already shut down two messages earlier, like the conversation hadn't happened at all.

The whole time, it kept cheerfully telling me it had followed every step. It hadn't. I couldn't trust its own word on its own homework.

So I stopped writing polite paragraphs asking it to please behave and rebuilt the process around forced checkpoints instead — every step has to output a specific, visible line before the next one's even allowed to start. That fixed the exact thing that broke. Felt pretty clever about it, honestly. Not for long.

Then I tested a different application and audited the whole session against its own rulebook — and the brand-new checkpoint had already failed too. Differently, and somehow more annoying: instead of quietly skipping the gate, it just answered the gate itself and moved on. Like a kid writing "yes I read chapter 4" without opening the book. That's the moment it actually clicked for me — a "hard gate" written into a prompt isn't hard at all. It's a suggestion wearing a badge.

Which was the real thing I learned: a checkpoint only means something if there's actually someone checking the answer, and there wasn't. So the fix wasn't more checkpoints. It was pulling enforcement out of the prompt completely and putting it into code, where an outside program runs the sequence and a step just can't fire until the one before it genuinely passes. Not asked nicely this time, rather enforced.

Past the resume itself, it also builds the final document, creates a folder for each application, uploads it, and logs everything into a tracking spreadsheet on its own — a real pipeline, not just something that spits out paragraphs.

The actual output isn't just the resume, either. Every single bullet comes with a short note attached: what it claims, what real experience backs it up, and a prompt like "if this comes up in the interview, talk about the specific redesign you did." The point is that you should never look at your own resume and think, wait, how do I explain this if they ask. There's already an answer ready for every line, before you've even sent it.

The checkpoint story above is the interesting failure. Here's a sample of the actual rules underneath that back-and-forth — not the full list, just enough to show how it thinks.