In “95% of People STILL Prompt ChatGPT-5 Wrong” by Jeff Su, published on September 23, 2025, the YouTuber delves into the surprising reasons behind the less effective output of ChatGPT-5 compared to its predecessors. According to Su, users often experience diminished results because OpenAI has fundamentally altered GPT-5’s architecture, rendering previous prompting techniques less effective. Despite GPT-5 being a more powerful model, the changes made necessitate new techniques to harness its capabilities effectively.
The first major update involves model consolidation; GPT-5 now functions with fewer models, and an invisible router decides which model handles a user’s request, a process not without its inefficiencies. This means that users who employ the same prompts as before might end up with suboptimal results if the router selects the least powerful model.
One of the well-supported arguments in Su’s presentation is the introduction of practical tips to interact with GPT-5. The “router nudge phrases,” for example, offer a strategy where specific trigger phrases like “think hard about this” force the AI to engage deeper reasoning, thereby improving output quality. This claim is effectively underpinned by demonstrating the difference in responses with and without these phrases, presenting a compelling case for this nuanced approach.
In addressing verbosity control, Jeff Su’s insights shine. He describes how users can adjust the output length to meet their unique needs by using simple phrases. This ability to modify verbosity allows users to tailor the AI output, ranging from concise messages to comprehensive documents, adding an element of flexibility that can enhance user experience. This section makes a compelling case for prompt customization and showcases the advanced adaptability of GPT-5.
However, while Su praises these improvements, he does not shy away from pointing out the challenges. Particularly, he notes that OpenAI’s emphasis on strict adherence to prompts can be a double-edged sword. With precise instructions, GPT-5 excels, but vaguer prompts no longer benefit from the AI’s previous ability to infer context. This highlights a limitation where relying on older, less structured prompting could yield poor results with GPT-5.
The introduction of the “perfection loop” exemplifies an advanced use case where the AI iteratively improves its outputs by grading and refining its responses. While Su presents this method as beneficial for tasks requiring high precision, it is also an acknowledgment of the complexity involved, which might be daunting for some users.
Ultimately, Jeff Su’s video not only educates viewers on optimizing GPT-5 usage but also provides thoughtful insights into adapting to evolving AI landscapes. His balanced critique acknowledges both the potential of new methodologies and the inherent challenges in transitioning to this advanced architecture, making this a valuable resource for those looking to adapt their approach to the latest AI technologies.