Prompt engineering is collapsing – GPT-5 just proved it
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.
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.
https://blog.big-picture.com/en/prompt-engineering-is-dead-i...