I’ve been developing something called the Universal Research Pass.
It’s not meant to be another “ultimate prompt.” It’s a structured research protocol designed to make ChatGPT investigate a question rather than simply answer it.
It tries to:
- challenge the starting assumption
- generate competing explanations
- distinguish discovery from proof
- trace source lineage
- investigate contradictions
- compare independent evidence
- calibrate confidence
- stop when there’s enough understanding to act
I’ve now frozen v1.0 as a public baseline and built a small open testing lab around it.
What I need is not praise. I need people to break it.
The basic test:
- Ask a real question normally in a fresh chat.
- In another fresh chat with the same model, load the Research Pass.
- Ask:
Research Pass this: [the exact same question]
- Compare the outputs.
I’m especially looking for:
- cases where the normal answer is better
- hallucinations
- over-research
- under-research
- failure to challenge a false premise
- poor source use
- unnecessary complexity
- model-specific weirdness
- modifications that improve one thing but break another
The project includes:
- canonical v1.0
- testing protocol
- structured failure reports
- improvement proposals
- contribution rules
- a small 10-case alpha benchmark
If you test it, please report what happened — especially if it failed.
GitHub repo: MagicalDealer/research-pass
Search that exact repo name on GitHub if you want to test it.
Welcome to the forum user.
I looked at it. It’s a long research playbook you feed ChatGPT, not an app and not really an “ultimate prompt” either—more like a checklist that tries to stop the model from just sounding sure. The bones are sensible: don’t protect the first belief, treat Reddit as a clue not a citation, look for rival explanations, say when you don’t know. What i actually think, after reading it, the idea is better than the artifact. v1.0 is frozen, documented, licensed, with a tester guide and a tiny 10-case benchmark, which is more honest than most prompt repos I look at. Whether it works is still the open question. My guess going in: on a messy real question it will sometimes surface a better frame; on a simple one it will write a term paper; and if the model doesn’t actually search, all the lineage/contradiction language just becomes a more elaborate hallucination.
This is pretty much exactly why I put it out publicly.
I don’t think v1.0 is the finished product, and I’m not particularly attached to any individual part of it. I feel like I came up with a strong skeleton, but I can only see the problems that occur in the kinds of research I personally do.
The experiment is basically: freeze the skeleton, give it to people with completely different backgrounds and use cases, and let them tear it apart hard enough that whatever survives becomes stronger.
Your three predictions are exactly the kind of thing I want to test — especially the “simple question turns into a term paper” problem and the possibility that a model without genuine research capability could turn the methodology into a more elaborate hallucination.
If you’re willing to actually run a few cases, even cases where it fails badly would be genuinely useful. I’m much more interested in finding the limits of the system than defending v1.0.
You should use more than one model, like test it on Gemini or another search capable UI. Then compare it straight no research pass vs with the structure. Gemini and co-pilot would be good choices bing vs google.
Yeah, I think you’re right. Cross-model testing needs to become a real part of the experiment rather than an occasional side test.
The important distinction I want to preserve is:
1. Normal vs. Research Pass within the same model, which tells us whether the protocol actually adds value.
2. Research Pass vs. Research Pass across models, which tells us how dependent the methodology is on the model/search system underneath it.
That also ties into your earlier point: if a model can’t genuinely retrieve or verify the evidence the protocol asks for, the structure might actually make a weak answer look more rigorous instead of making it more correct.
I’m going to add cross-model behavior as something we explicitly track. Gemini and Copilot make sense as early comparisons alongside ChatGPT.