Prompt engineering is collapsing – GPT-5 just proved it | Hacker News Reader

Prompt engineering is collapsing – GPT-5 just proved it

10 pointsyuer20258 comments

GPT-5 is a beast. But here’s the thing nobody wants to say out loud: it just killed prompt engineering as a sustainable practice.

Carefully tuned prompts from GPT-4o? Broken.

Styles, logic, answer habits? All shifted.

Companies? Forced to roll back or re-test thousands of prompts overnight.

This isn’t progress. It’s technical debt disguised as innovation. Every new release means paying a Prompt Migration Tax: rewriting, regression-testing, and re-training teams.

Meanwhile:

Users are losing trust — sticking with old models or switching providers.

Security is a joke — OWASP already flagged prompt injection as the #1 LLM risk, and NIST said the same.

Vendors keep pushing “best practices” like longer separators or system prompts… band-aids on a structural wound.

The cycle looks like this: upgrade → break → patch → break again → patch again. How long before the entire industry realizes this is a dead end?

Prompt engineering isn’t the future. It’s a trap. And GPT-5 just made that painfully clear.

Not sure why people are so quick to disagree with this. The premise is roughly true. If we build prompt architectures (which can be massive) on a given model (say gpt4o) and then that model is just removed/no longer available, the entire architecture must be reworked for the newest model. What if Python changed its syntax every 6 months to a year and all Python code in production would only work if it was updated to the latest Python? That’s the issue… even if we maintain access to and use legacy models, they might be super expensive or gpu inefficient or slow, etc compared to new models. It’s very tumultuous ground to construct on.

Just use DPSy.

Agreed this is why compound systems are the future of AI.

https://youtu.be/vRTcE19M-KE

This looks AI written. It's full of AI writings tropes and, the big telltale sign: it's a lot of words for saying very little.

Obviously, you have to rewrite prompts for different models.

If you are really dependent on a single one; then better be sure it's an open-source copy you can run yourself.

This isn’t true for 99% of the prompts that actual businesses are engineering and using. Your typical user is clueless and just needs a prompt that “translates” their ad hoc, often ambiguous, questions into clearly specified tasks that the AI understands without ambiguity and states the underlying assumptions the user was taking for granted. Such prompts generalize well to any language model. You’re foolish and a bad prompter engineer if you engineer niche prompts that don’t generalize well and are specific to just one model—you can engineer a prompt that generalize and accomplishes the same thing. Prompt Engineering isn’t dead—we just have a new “lesson learned” for the previously naive. Learn it!

Prompt engineering died like a year ago

https://blog.big-picture.com/en/prompt-engineering-is-dead-i...

didnt prove anything. prompt engineering still works with GPT-5. dont know what your experience is about...

I completely disagree. Just signed up for Perplexity Pro and don't trust anything it says, but lots and lots of ideas I didn't think of.

Sorta like RT and Sydney Morning News. Gives me leads, not necessarily truth (and good golly miss molly, a lot of Australian news is about sports!)

Reply on news.ycombinator.com