# 5 Prompting Tricks to Make Your AI Less Average Source: https://www.youtube.com/watch?v=JdXlQeQcFN4 Recap page: https://rapidrecap.app/video/JdXlQeQcFN4 Generated: 2025-11-04T13:08:55.6+00:00 --- ## Quick Overview To make a Large Language Model (LLM) less average and achieve high-quality, distinct output, users must move beyond simple prompting and implement strategies like using a "Negative Style Guide" to explicitly tell the model what to avoid, such as common clichés, telemetric language, or specific formatting like colons and dashes, as demonstrated by the speaker's successful use of these techniques with both GPT-5 and O3 models. **Key Points:** - LLMs trained on the entire corpus of human output suffer from an AI sameness problem, producing average outputs. - The speaker advocates for a "Negative Style Guide" as a core technique to make LLM output non-average and distinct. - This technique involves explicitly instructing the LLM what to avoid, such as clichés, telemetric language, or specific formatting like colons and dashes. - The speaker successfully used this method with both GPT-5 and O3 models to generate superior outputs compared to standard prompting. - A key goal is to force the model to choose between divergent paths rather than defaulting to the average consensus, exemplified by asking it to argue for one option over another. - The speaker notes that the O3 model struggled to avoid certain patterns, requiring constant reminders, while the newer GPT-5 thinking was better at incorporating these constraints. - A related technique is "Self Critique," where the model is asked to critique its own output using examples to explain why the consensus view is wrong. ![Screenshot at 0:05: The speaker introduces the core concept of overcoming the 'tyranny of the average' in LLM outputs by employing specific prompting strategies.](https://ss.rapidrecap.app/screens/JdXlQeQcFN4/00-00-05.png) **Context:** The video addresses the common issue where AI-generated content, having been trained on the entire human corpus, often defaults to producing average, uniform, or cliché-ridden output. The speaker frames this as AI's "tyranny of the average," arguing that achieving truly unique and high-quality results requires intentional deviation from this baseline. The discussion focuses on practical prompting strategies developed through the speaker's personal experience using models like GPT-5 and O3. ## Detailed Analysis The video explains that because Large Language Models (LLMs) are trained on the average of all human-generated content, their default output tends toward the mean, resulting in uninspired or cliché writing. To combat this 'AI sameness problem,' the speaker introduces five prompting tricks, centered around the concept of a "Negative Style Guide." This guide involves explicitly telling the AI what *not* to do, such as forbidding telemetric language, colons, dashes, or specific formatting, and forcing it to argue for one choice over others to avoid settling on the statistical average. The speaker references an article by Alex Kantrowitz titled 'AI's Sameness Problem' to support the necessity of breaking this uniformity. The speaker successfully used these techniques with both GPT-5 and O3 models, noting that O3 often required explicit reminders to avoid common pitfalls like citing the same tired examples, whereas GPT-5 seemed better at internalizing the constraints. The final technique discussed is Self Critique, where the model is prompted to generate examples explaining why conventional consensus is wrong, forcing a deeper level of reasoning and originality in the output. ### The Sameness Problem - AI output averages averages - tends to produce output optimized around the mean of human-generated work - novelty wears off fast. ### Technique 1 - Negative Style Guide: Explicitly list forbidden elements (e.g., telemetric language, colons, dashes) to force distinction - requires constant reminding for older models like O3. ### Technique 2 - Forcing Choice: Prompt the model to argue for one option over others, preventing it from defaulting to the consensus or average argument. ### Technique 3 - Self Critique: Instruct the model to critique its own output, specifically using examples to explain why the general consensus is wrong. ### Application and Results - These techniques yield higher quality, more distinct outputs compared to standard prompting across models like GPT-5 and O3. ![Screenshot at 0:05: The speaker introduces the core concept of overcoming the 'tyranny of the average' in LLM outputs by employing specific prompting strategies.](https://ss.rapidrecap.app/screens/JdXlQeQcFN4/00-00-05.png) ![Screenshot at 0:37: Visual representation of the output being based on how the user prompted it, setting up the need for better prompting.](https://ss.rapidrecap.app/screens/JdXlQeQcFN4/00-00-37.png) ![Screenshot at 1:16: Reference to Alex Kantrowitz's essay on the 'AI's Sameness Problem,' highlighting the core issue being addressed.](https://ss.rapidrecap.app/screens/JdXlQeQcFN4/00-01-16.png) ![Screenshot at 2:36: The speaker begins to explain the first part of the solution: avoiding specific formatting like colons and dashes.](https://ss.rapidrecap.app/screens/JdXlQeQcFN4/00-02-36.png) ![Screenshot at 4:36: A visual metaphor of a teacher instructing a robot student, symbolizing the process of refining AI output with specific guidance \(Negative Style Guide\).](https://ss.rapidrecap.app/screens/JdXlQeQcFN4/00-04-36.png) ![Screenshot at 6:43: A switch to a different classroom scene illustrating a concept titled 'Force Divergence and Choice,' relating to forcing the LLM to choose a path.](https://ss.rapidrecap.app/screens/JdXlQeQcFN4/00-06-43.png) ![Screenshot at 9:09: A transition to a second classroom scene where the teacher points to a blackboard listing 'CLICHE BURN DOWN,' symbolizing the negative style guide technique.](https://ss.rapidrecap.app/screens/JdXlQeQcFN4/00-09-09.png) ![Screenshot at 11:11: A third classroom scene appears, showing a different robot model and the concept 'Switch Models' on the board, illustrating architectural approaches.](https://ss.rapidrecap.app/screens/JdXlQeQcFN4/00-11-11.png) ![Screenshot at 13:15: The final classroom scene shows the metallic robot model being instructed to 'Use Examples \(and explain why\) the consensus is wrong,' representing the Self Critique technique.](https://ss.rapidrecap.app/screens/JdXlQeQcFN4/00-13-15.png)