We ve spent the last few months moving our Google Workspace tools over to the Model Context Protocol (MCP). While the potential for 'agentic' workflows is huge, but the friction of working on the edge of a new protocol is very real.

For those who haven't dived in yet, the biggest hurdle we found was the Remote vs. Local setup, rather than the protocol itself. Most tutorials focus on local command-line installs, but for a production-ready SaaS, you have to build for remote MCP servers. This requires a completely different approach to authentication and persistent state that the current docs don't fully cover yet.

180 applications for one senior engineering role last month. About 40 read like the same resume.
Same action verbs. Same keyword density mapped to the JD. Same metrics with no context attached. Written for the parser, not for a person.
The standard fix is detection. Spot the AI resume, discount it. But detection only tells you the document is unreliable. It tells you nothing about who is actually good. You have filtered noise and gained zero signal.
So the first filter is now doing no work, and everyone is still running it.
I have started skipping the resume as the opening screen and looking at what someone has actually shipped instead. Better signal. Slower. Does not scale cleanly yet, and it disadvantages people whose best work is behind an NDA.
Disclosure, I build hiring software, so I am biased toward thinking the resume is the wrong input. Which is why I want the argument and not the agreement.
Two questions.
  What phrase are you seeing on repeat? Mine is spearheaded cross-functional initiatives.
  If you dropped the resume as filter one, what replaced it, and what broke?

I'm working in AI text detection, and the honest output of that kind of model is a probability, not a verdict. Users don't want a probability. Hand someone 78% and they read it as a yes, then they treat the 22% as if it were never there.

That gap between what a model knows and what the screen says has been the hardest part of the build for me, and it isn't specific to detection. Anything with a model behind it has the same gap: a spam score, a risk score, a match score, a credit score, a health ring.

How much of that uncertainty do you actually put on screen?

Here's where I've landed so far, and I'm not confident about any of it:

Today, I learned that OpenAI has changed their policy on sharing custom GPTs. In the past, one could create a GPT using their platform that you could then share with whomever you wanted, however you wanted.
Now, with any NEW GPTs you make, you can no longer share if you're on a "non-business" plan.
Here's their policy: https://help.openai.com/en/artic... Look under "Availability."
I have my own thoughts on this (mixed). What do you think?

Anything generative has a failure rate you can't fully engineer away. Mine is low but it isn't zero, and when it misses, the user has already spent one of a small number of credits.

I've gone back and forth on this and landed somewhere I'm still not sure about.

What I considered:

1. Free retries. Cleanest for the user and the one I wanted. The problem is a retry costs me exactly what the original cost, and "free" has no natural stopping point someone who doesn't like the output can re-roll indefinitely, and I can't tell dissatisfaction apart from someone farming variations.

2. One free retry per generation. Fairer, but it turns every borderline result into a judgement about whether it "counts" as a failure, and I'd be the one judging.

3. Charge for it, and price credits assuming a share of them get spent on retries. This is what I do now. It's honest about the compute, but it means a bad result costs the user something on the day it happens to them, which is the worst possible day for that.

4. Refund on request, handled in support. No policy, no automation. Doesn't scale, but might be correct at my size.

and how i got hunters to back my launch

most people treat product hunt like a slot machine. pull the lever, hope for upvotes. but if you re launching anything serious, that s a waste.