95% of People STILL Prompt ChatGPT-5 Wrong
About this video
Jeff Su argues that after GPT-5’s launch, users got worse results not because the model is weaker but because OpenAI changed its architecture, and he says he spent a month testing GPT-5 with OpenAI’s official guides. Two big changes: (1) model consolidation — Plus users went from many models to three (GPT-5, GPT-5 Thinking Mini, and GPT-5 Thinking), with an invisible router deciding which model handles each request, sometimes picking the fastest rather than the smartest; (2) GPT-5 follows instructions with extreme precision (it was trained with AI agents in mind) but is worse at guessing intent from vague prompts. He gives five fixes, easiest to hardest. Tip 1, router nudge phrases: adding ’think hard about this’ (also ’think deeply about this’ or ’think carefully’) forces deeper reasoning, while ’this is critical’ and ’this is very important’ did not work in his testing. Tip 2, verbosity control: use phrases like ‘give me the bottom line in 100 words or less’ for short outputs, ‘aim for a concise 3 to five paragraph explanation’ for medium, and ‘provide a comprehensive and detailed breakdown, 600 to 800 words’ for long ones. Tip 3, OpenAI’s official prompt optimizer rewrites prompts by adding structure, eliminating vagueness, and adding error handling; since it needs a separate developer account and payment method, he offers a free workaround meta-prompt that asks GPT-5 itself to act as an expert prompt engineer and improve your prompt. Tip 4, the ‘XML sandwich’: wrap prompt sections in labeled XML tags (task, resume, job description, tone) so the model can distinguish background, task, and output format. Tip 5, the perfection loop: instruct GPT-5 upfront to build an internal rubric for excellence, grade its own draft, and iterate internally until it scores top marks — best for complex zero-to-one tasks like finished documents or production-ready code. The tips stack and can be combined.
Key takeaways
- GPT-5’s invisible router may route a plain prompt to a weaker model; ending with ’think hard about this’ reliably triggers deeper reasoning
- Control output length explicitly with verbosity phrases, from ’the bottom line in 100 words or less’ up to a ‘comprehensive and detailed breakdown, 600 to 800 words’
- Use OpenAI’s prompt optimizer — or a free meta-prompt asking GPT-5 to act as an expert prompt engineer — to add structure, remove vagueness, and add error handling
- Structure prompts as an ‘XML sandwich’ with labeled tags for task, context/background material, and output format or tone
- For complex tasks, use the perfection loop: have GPT-5 create its own rubric, grade its draft, and iterate internally until it scores top marks
Episode details: published September 23, 2025 · 11:51 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are Jeff Su’s own claims, shared for education — they are not typical or guaranteed.
← Previous: My Top 5 AI Tips for iPhone Users! · Next: How to Create Cinematic AI Videos (No-BS Guide) →
All Jeff Su videos · Jeff Su profile · Jeff Su (official site)
