95% of People STILL Prompt ChatGPT-5 Wrong

About this video

• Video Title: 95% of People STILL Prompt ChatGPT-5 Wrong
• Channel: Jeff Su
• Speakers: Jeff Su
• Duration: 00:11:51

Overview

This video explains why users are experiencing worse results with ChatGPT-5 despite it being a more powerful model. The presenter, Jeff Su, details two fundamental architectural changes in GPT-5 by OpenAI: model consolidation with an invisible router and improved instruction following. He then provides five actionable tips to adapt prompting techniques for GPT-5 to achieve better outputs.

Key takeaways

• Architectural Changes in GPT-5: OpenAI has consolidated models into three main versions (GPT-5, GPT-5 thinking mini, GPT-5 thinking) and introduced an invisible router that selects the model for a request. This router can sometimes lead to less optimal model selection. Additionally, GPT-5 is significantly better at following instructions, which is a double-edged sword: it's precise but less forgiving of vague prompts.
• Tip 1: Router Nudge Phrases: Adding specific phrases like "Think hard about this" can influence the router to select a higher reasoning model, leading to more nuanced and comprehensive answers with consideration for second-order effects.
• Tip 2: Verbosity Control: Users can control the length and detail of GPT-5's output by using specific phrases such as "give me the bottom line in 100 words or less" for concise outputs, "aim for concise 3 to five paragraph explanation" for medium detail, and "Provide a comprehensive and detailed breakdown" for extensive outputs.
• Tip 3: OpenAI's Prompt Optimizer: This tool (accessible via a developer account) can rewrite prompts for GPT-5, adding structure, eliminating vagueness, and including error handling. A free workaround involves using a meta-prompt that instructs GPT-5 to act as a prompt engineer and improve the given prompt.
• Tip 4: XML Sandwich Structure: Organizing prompts using XML tags (e.g., <context>, <task>, <output_format>) helps GPT-5 better understand and follow instructions by clearly labeling different components of the prompt. This is particularly useful for custom instructions and GPTs.
• Tip 5: The Perfection Loop: This advanced technique involves instructing GPT-5 to develop its own rubric for excellence, internally iterate on its response based on that rubric, and refine it until it meets the criteria. This is best for complex, zero-to-one tasks.