Summary
Highlights
The video introduces a GPT-5 prompting trick called self-reflection prompting, which involves the AI creating a rubric, critiquing its own work, and redoing it internally to produce an improved output. This technique, validated by OpenAI, can lead to a 5-40% improvement in quality across various tasks like math, reasoning, analysis, and coding, while also decreasing hallucination rates.
The self-reflection process begins with the prompt, leading GPT-5 to plan by creating a private rubric with five to seven internal categories. It drafts an initial response, critiques it against the rubric, and iteratively redoes the work until all categories are met with high accuracy (e.g., 90%). This internal, iterative process (5-7 times) ensures a refined, high-quality final output without exposing intermediate drafts to the user.
The base prompt instructs GPT-5 to create a private rubric of 5-7 items relevant to task excellence, drafted iteratively. The 5-7 item cap prevents overthinking and degradation of intelligence. The prompt emphasizes keeping the critique internal, ensuring all tokens are used for reasoning rather than output, leading to more effective AI intelligence.
An optional section allows for the creation of alternative drafts in critical tasks, letting the AI choose the best iterated version. Stopping criteria are included to prevent the AI from overthinking and degrading its intelligence once the rubric criteria are passed.
For research tasks, the core prompt structure remains consistent. The task definition dictates the rubric categories. Explicit critiques can be added, such as ensuring at least three claims are backed by credible sources. The output can be formatted into specific sections like 'What's New,' 'Risk,' and 'Next Steps,' with a user-defined rubric for excellence (e.g., accuracy, recency, completeness, clarity).
In writing tasks (emails, blogs, LinkedIn posts), the voice of the writing dictates the rubric's categories. An explicit critique can be added, such as ensuring a strong hook that is concrete, beneficial, and cliché-free. The prompt defines the desired output. The rubric can specifically address elements like hook strength, specificity, structure, brevity, tone, and skimability.
For analysis tasks, the goal of the analysis drives the rubric categories. An additional requirement can be to include the top three assumptions with confidence levels in the output. A specific critique might be to address at least one strong counterargument internally. The rubric can include elements like decision rationale, risk assessment, and clear next steps.
For low-stakes activities, a simple rubric might suffice, allowing the AI to define categories. For high-risk tasks (public, factual, code-oriented), being explicit about rubric items and adding one to three explicit self-critiques is recommended. Key takeaways include capping rubric items at 5-7 to prevent overthinking, showing only the final output to optimize token usage, and using alternate options for high-stakes situations.