TLDR;
This video discusses how to effectively prompt GPT-5 to achieve optimal results, arguing that the model requires a different prompting approach than other AI models. It covers key techniques such as controlling reasoning level and verbosity, leveraging tool calling, utilizing self-reflection with rubrics, and employing metaprompting to refine prompts. The video also introduces the OpenAI prompt optimizer, a tool designed to enhance prompts specifically for GPT-5.
- GPT-5 requires specific prompting techniques due to its nature as a router model.
- Controlling reasoning level and verbosity is crucial for desired outputs.
- Tool calling, self-reflection, and metaprompting are essential for advanced results.
- The OpenAI prompt optimizer can significantly improve prompt effectiveness.
Intro [0:00]
The video addresses the initial negative reactions to GPT-5, suggesting that the model's true potential is unlocked through proper prompting techniques. The presenter claims that using specific methods can yield results superior to those of other AI models. The video promises to reveal four prompting methods and a free tool to enhance prompt quality, ultimately granting users "AI superpowers" with GPT-5.
Reasoning and verbosity [1:01]
GPT-5 differs from other models as it is a router model, requiring users to manually control variables like reasoning level and verbosity. Reasoning level determines the amount of time the model spends thinking, while verbosity dictates the depth of the response. The presenter emphasizes the importance of specifying these variables in each prompt to avoid generic outputs. For instance, when creating a Discord community, the presenter initially received bland answers until specifying a higher reasoning level ("think harder"), which resulted in detailed channel suggestions, roles, permissions, and even emoji recommendations. Verbosity is controlled through levels (low, medium, high) to determine the depth of the answers.
Tool calling [6:09]
GPT-5 has access to more tools than other AI models and can use multiple tools within a single prompt. To illustrate this, the presenter uses the prompt builder within chat GPT and writes a prompt for community planning that leverages multiple tools. The prompt requests a logo, a brand guideline PDF, a community announcement tweet, and web research on competitors, also specifying "think hard" and "medium verbose answer". GPT-5 successfully generates a logo, a brand guideline PDF, tweets, and competitor research, demonstrating its ability to use multiple tools in one prompt.
Self reflection [10:12]
GPT-5 is adept at self-reflection, which can be used to improve results, especially in coding. The presenter uses a prompt to build a first-person shooter game using 3JS, incorporating a self-reflection component. This involves instructing GPT-5 to create a grading rubric for itself, think deeply about what makes a world-class app, and iterate on the solution until it meets the rubric's standards. The resulting game features detailed enemies, an environment with a ring, and sound effects, showcasing the power of self-reflection in enhancing code quality.
Metaprompting [15:30]
Metaprompting involves having the AI refine its own prompts to improve results. If an initial prompt yields unsatisfactory results, metaprompting can be used to identify and address shortcomings. The presenter revisits the game example, specifying the desired behavior (more complex enemies and power-ups) and the undesired behavior (their absence in the game). GPT-5 then suggests prompt adjustments, such as explicitly asking for enemy archetypes with distinct behaviors and collectible power-ups with effects and timers.
Prompt Optimizer [17:57]
The OpenAI prompt optimizer is a tool that refines prompts for specific models, including GPT-5, by leveraging its knowledge of the model's strengths and weaknesses. The presenter inputs the original game generation prompt into the optimizer, which then outputs a more detailed and effective prompt. The optimized prompt includes instructions to perform an in-depth self-reflection and construct a comprehensive private rubric. The presenter emphasizes that using the prompt optimizer can significantly improve results.