TLDR;
This video presents a comprehensive prompting system for GPT-5, designed to significantly improve the quality of its outputs. It emphasizes a five-part master prompt that includes defining the AI's role, setting up a control panel to manage parameters like reasoning and verbosity, specifying a clear task, providing relevant inputs or context, and detailing the desired deliverables. The system also incorporates "private ops" to guide GPT-5's self-reflection and metapixing processes, which are crucial for refining its responses.
- Role definition for specific task execution.
- Control panel for managing reasoning, verbosity, and tool usage.
- Task specification for clear objectives.
- Input context for relevant information.
- Deliverables for precise output requirements.
Intro [0:00]
The video introduces a master prompt for GPT-5, claiming it can significantly improve results by guiding the AI through five key stages. It addresses the common issue of GPT-5 producing subpar outputs when not prompted correctly. The master prompt is based on OpenAI's cookbook and aims to consolidate hours of advice into a simple, effective template.
Role [1:31]
The first part of the master prompt involves defining the role GPT-5 should assume. This is crucial because GPT-5 strictly adheres to instructions, doing neither more nor less than what it's told. For example, in the demonstration of planning an app for YouTube creators, the role is specified as "a product full stack app planner for indie creators." The role should be tailored to the specific task to ensure the AI understands its function from the outset.
Control Panel [2:46]
The second part of the master prompt is the control panel, which allows users to manage various parameters of GPT-5, such as reasoning, verbosity, and tool usage. This section covers reasoning (think, think hard, or ultra think), verbosity (low, medium, or high), and tools (web, code, PDF, images, or auto). Two key features to control are self-reflection (reflecting before execution to improve the prompt) and metaphix (reflecting after execution to refine the results). These features enhance GPT-5's ability to deliver high-quality outputs, although they may increase processing time.
Task [6:28]
The third part of the master prompt is a simple, one-sentence task that tells the model what to do. For the app-planning example, the task is to "plan and scaffold a minimal app called YouTube topic scout that finds trending ideas, scores them, and generates a script outline." The task should be concise, as the other parts of the prompt provide the necessary details and context.
Inputs [7:30]
The fourth part of the master prompt involves providing inputs, which is optional but valuable for adding context to the task. Inputs can include notes, links, data, ideas, and comparisons. For the app-planning example, inputs include user information, core loop descriptions, scoring preferences, non-negotiables, tech preferences, and tone. This section is about "brain dumping" as much relevant information as possible to guide the AI.
Deliverables [9:24]
The fifth part of the master prompt specifies the deliverables, which are the precise outputs desired from GPT-5. In the app-planning example, deliverables include a product requirements document (PRD), competitor scan, architecture, API spec, UI design, and starter code. By listing these deliverables, the user directs GPT-5 to produce a comprehensive set of results.
Results [14:26]
The video concludes by demonstrating the results of using the master prompt. GPT-5 generated code for the app, a PRD, a competitor scan, architecture details, an API spec, wireframes, and a zip file with all the code. The AI even applied metaphix to improve the PRD based on a self-scoring rubric. The presenter emphasizes that while this detailed prompt is most useful for initiating complex tasks, subsequent prompts in the same conversation can be simpler, leveraging the context already established.