Maximizing GPT-5’s Potential: Andrew Bolis on Tailoring Prompts for Advanced Reasoning

In a recent LinkedIn post, Andrew Bolis highlights a common pitfall in utilizing advanced AI models like GPT-5, suggesting that many users interact with it as if it were an earlier version, thereby missing out on its enhanced reasoning capabilities. Bolis emphasizes that unlocking GPT-5’s full potential requires a strategic approach to prompting, akin to how one would brief a human consultant based on the complexity of the task.

Understanding GPT-5’s Built-In Reasoning Modes

Bolis introduces the concept that GPT-5 possesses distinct built-in reasoning modes, which users need to actively engage. He explains that the effectiveness of the AI’s output is directly tied to how well the prompt guides its thinking process. “To get the most out of GPT-5, you need to control how it thinks,” Bolis asserts, drawing a parallel to differentiating instructions for a quick update versus a comprehensive strategic plan.

“Most people prompt GPT-5 like it’s GPT-4. And miss out on its new reasoning power.”

This core observation sets the stage for Bolis’s detailed breakdown of how to leverage these modes for optimal results.

Leveraging High and Minimal Reasoning Modes

For tasks demanding deep analysis, strategic planning, or innovative problem-solving, Bolis recommends employing what he terms ‘High Reasoning Mode’. He provides specific prompt directives for this mode, such as instructing the AI to “Think comprehensively before answering” or to “Reason more systematically.” An example prompt he shares for this mode is:

“You are my business advisor. Reason more systematically. Draft a 1-page strategy for [organization] to achieve [objective] within [timeline].”

Conversely, for scenarios requiring swift responses, brief overviews, or simple task execution, Bolis advocates for ‘Minimal Reasoning Mode’. This approach involves using prompts that explicitly state the need for brevity and directness, like “Minimal reasoning needed” or “Provide solution immediately.” He uses this mode daily for generating concise content, as demonstrated by his prompt example:

“Minimal reasoning. Create 5 email subject lines for [service] targeting [audience]. No explanations required.”

Exploring Chain-of-Thought and Tree-of-Thought

Beyond these two primary modes, Bolis delves into more sophisticated reasoning techniques. ‘Chain-of-Thought’ (CoT) prompting is presented as ideal for tasks that benefit from a sequential, step-by-step approach. Bolis suggests using phrases like: “Apply step-by-step reasoning: divide the challenge into stages, address each, then provide the final solution.” This method is particularly useful for complex business challenges, performance reviews, and content planning.

For situations requiring the evaluation of multiple possibilities, Bolis introduces ‘Tree-of-Thought’ (ToT) prompting. This technique encourages the AI to explore various pathways, weigh pros and cons, and ultimately recommend the best option. He advises prompts such as: “Generate at least 3 options. Weigh the advantages and drawbacks of each, then recommend the strongest choice and explain why.” This is presented as highly effective for strategy development, investment decisions, and generating creative solutions.

The Complete Optimization Framework

To synthesize these techniques, Andrew Bolis outlines a comprehensive five-step framework for optimizing AI interactions:

1. Define the role (e.g., “You are a…”).
2. Provide background information (objective, audience, limitations).
3. Set the reasoning approach (high, minimal, CoT, or ToT).
4. Outline the desired format for the output.
5. Request iterative improvements or critical review.

Bolis concludes by reiterating that many users are underutilizing GPT-5’s advanced capabilities. “Your prompt should guide the work, not guess it,” he stresses, advocating for a deliberate match between the task’s complexity and the chosen reasoning level. By understanding and applying these distinct prompting strategies, users can significantly enhance the utility and accuracy of their AI interactions.