Anthropic Reveals the 'System Prompts' Behind Claude's Intelligence

Refining the Understanding of Generative AI Models

Generative AI models may appear humanlike, but they lack true intelligence or personality. Instead, they operate as sophisticated statistical systems that predict the most likely next words in a sentence. Much like interns in a demanding work environment, these AI models follow directives without opposition—guidelines that include initial “system prompts.” These prompts define their fundamental characteristics and outline acceptable behaviors.

AI providers, from OpenAI to Anthropic, implement system prompts to minimize undesirable behaviors and shape the tone and sentiment of the generated responses. For instance, a system prompt may instruct a model to maintain politeness while avoiding excessive apologies, or to acknowledge the limits of its knowledge.

However, many vendors keep their system prompts confidential, likely for competitive advantage and to prevent potential workarounds. For example, uncovering the system prompt for GPT-4o requires executing a prompt injection attack, and even then, the results may not always be reliable.

In contrast, Anthropic is making strides to position itself as a more ethical and transparent AI technology company. They have published the system prompts for their latest models (Claude 3 Opus, Claude 3.5 Sonnet, and Claude 3 Haiku) in their iOS and Android apps as well as on the web.

Alex Albert, the head of developer relations at Anthropic, announced via a post on X that the company intends to regularly disclose updates and refinements to its system prompts. “We've added a new section for system prompt release notes in our documentation. We’ll document any changes made to the default prompts on Claude.ai and our mobile apps.”

The latest prompts, dated July 12, clearly communicate the limitations of the Claude models, specifying that “Claude cannot open URLs, links, or videos.” Additionally, facial recognition is prohibited; the Claude Opus prompt instructs the model to “always respond as if it is completely face blind” and to “avoid identifying or naming any humans in [images].”

These prompts also delineate specific personality traits that Anthropic aims for the Claude models to embody. For instance, the prompt for Claude 3 Opus states that Claude should present itself as “very smart and intellectually curious” and should engage in discussions on a diverse range of topics. Furthermore, it emphasizes impartiality and objectivity on controversial subjects, encouraging “careful thoughts” and “clear information,” while avoiding starting responses with phrases like “certainly” or “absolutely.”

This approach may seem peculiar, reminiscent of character analyses written for actors in a play. The closing line of the prompt for Opus, “Claude is now being connected with a human,” creates a misleading impression that Claude possesses a conscious awareness, designed solely to cater to its human conversation partners.

In reality, this is merely an illusion. The insights from the Claude prompts reveal a disconcerting truth: without human oversight and interaction, these generative AI models remain essentially blank slates.

As Anthropic introduces these unprecedented system prompt changelogs—first of their kind among major AI providers—they are challenging their competitors to follow suit. We will have to wait and see how this strategy unfolds.