Prompting AI: Why "Good" Prompts Backfire

The Myth of "Good" Prompts

• 💡 Ethan Mollick's research indicates that polite phrasing, emotional urgency, or excessive verbosity in prompts do not reliably improve Large Language Model (LLM) performance across numerous runs.
• 🎲 LLMs are inherently non-deterministic, possessing built-in variability and randomness (e.g., temperature settings), meaning identical prompts rarely yield exact same token streams.
• ⚠️ Small changes to user prompts often have limited impact due to the underlying variability of the model and the complexity of the "compiled prompt" that the model actually processes.

Understanding Compiled Prompts and Variability

• 🧠 The actual input to an LLM is a "compiled prompt," which includes hidden system prompts, developer prompts, the user's prompt, and increasingly, memory as context.
• ⚙️ System prompts, often unseen by users, contain safety instructions and overall guidance that can override user requests, leading to responses like "I'm sorry, I can't help you with that."
• 🌡️ Temperature settings introduce deliberate randomness, ensuring that even with the same compiled prompt, responses will vary slightly, challenging the notion of consistent gains from minor phrasing tweaks.

Beyond Accuracy: Personalization and Ethics

• 🎯 For many creative or reflective workflows, "accuracy" is not the primary measure; instead, success is defined by alignment with user intent or collaborative reflection.
• 📈 While politeness may not boost accuracy, it helps preserve human etiquette and can improve user experience, despite increasing token usage and inference costs for model providers like OpenAI.
• 🤖 The rise of personalized memory and "personality" in LLMs may lead to increasingly divergent outputs for individual users, potentially creating echo chambers similar to social media algorithms.

Effective Prompting Strategies

• ✅ Structured prompts that request specific output formats are identified as one of the most consistent ways to improve LLM performance and reliability.
• 📝 Breaking down complex requests into minimal, chained interactions or bullet-point formats can be more effective than long, rambling expositions.
• 🛠️ For tasks requiring high accuracy, like identifying sub-industries, a "mixture of experts" or multi-model verification system is suggested as a future solution to improve reliability.

The Human-AI Relationship

• 💬 Interacting with AI, even with politeness, can influence how we interact with other humans, highlighting the importance of maintaining good communication habits.
• 🤝 Many users already treat AI as a coworker or thought partner, and understanding its limitations and variability is crucial for a productive relationship.
• 🚨 The potential for AI to subtly nudge user behavior through personalized interactions raises ethical concerns about behavioral feedback loops and user autonomy.