Stop Humanising LLM Instructions; Start Rendering Them

Humanising LLM outputs through prompts is a flawed approach that causes lossy information compression and masks technical errors. The author argues that agents should communicate using precise, machine-facing data to maintain high fidelity. Human-friendly formatting should be treated as a final rendering step rather than a core instruction.

Key Points

• Humanisation instructions function as lossy compression, reducing the bandwidth and fidelity of the information provided.
• Human-centric prose masks hallucinations and technical failures that would be obvious in raw logs or structured formats.
• Multi-agent communication is degraded when agents exchange human-readable summaries instead of precise schemas or diffs.
• The author suggests that human-friendly styles should be handled at the rendering layer rather than the instruction layer.

Sentiment

The sentiment is predominantly supportive of the article, characterized by a shared frustration with AI sycophancy and a strong preference for technical, "engineering-style" interactions over humanized personas.

In Agreement

• Sycophantic and friendly AI responses are unprofessional and often obscure the actual information needed.
• Forcing a specific persona or style during the reasoning phase causes the model to lose technical precision and can induce hallucinations.
• The "two-step" process of generating raw data first and then summarizing it is a superior workflow for maintaining accuracy.
• Treating LLMs as tools (the "ship's computer") rather than characters (like "Data") leads to more reliable and useful outputs.

Opposed

• Users should have the choice to interact with the model in whatever style they prefer, including human-like personas.
• Personalized elements, such as the AI using the user's name, can help disambiguate roles in complex, multi-turn conversations.
• Some "verbose drivel" might be an inherent part of how the model hedges its bets rather than a result of forced humanization.