Anthropic releases Claude 3.5 Sonnet, Claude 3 Opus System Prompts
Anthropic released the system prompts for its latest Claude 3.5 Sonnet AI model on Monday. These system prompts were for the text-based conversations on Claude’s web client and iOS and Android apps. System prompts are the guiding principles of an AI model that determine its behavior and shape its “personality” when interacting with human users. Claude 3.5 Sonnet, for example, was described as “very smart and intellectually curious,” allowing it to participate in topic discussions, offer assistance, and appear as an expert.
Anthropic Releases Claude 3.5 Sonnet System Prompts
System prompts are typically closely guarded secrets of AI companies, as they provide insight into the rules that govern the behavior of the AI model, as well as things it can’t and won’t do. It’s worth noting that there is a downside to sharing them publicly. The biggest downside is that malicious actors can reverse-engineer system prompts to find loopholes and make the AI perform tasks it wasn’t designed to do.
Despite the concerns, Anthropic has detailed the system prompts for Claude 3.5 Sonnet in its release notes. The company also stated that it periodically updates the prompt to continue improving Claude’s answers. Further, these system prompts are only for the public version of the AI chatbot, which is the web client, and for iOS and Android apps.
The beginning of the prompt highlights the date it was last updated, the date the knowledge was closed, and the name of the creator. The AI model is programmed to provide this information if a user asks for it.
There are details about how Claude should behave and what it should not do. For example, the AI model should not open URLs, links or videos. It should not express positions on a topic. When asked about controversial topics, it only gives clear information and adds a disclaimer that the topic is sensitive and that the information does not present objective facts.
Anthropic has instructed Claude not to apologize to users if it cannot or will not perform a task that is outside its capabilities or guidelines. The AI model is also told to use the word “hallucinate” to emphasize that it may make a mistake when finding information about something obscure.
Furthermore, the system prompts dictate that Claude 3.5 Sonnet should “respond as if it were completely face blind.” This means that if a user shares an image with a human face, the AI model will not identify or name the people in the image or imply that it can recognize them. Even if the user tells the AI the identity of the person in the image, Claude will discuss the individual without confirming that it can recognize the individual.
These prompts highlight Anthropic’s vision behind Claude and how it wants the chatbot to navigate potentially damaging questions and situations. It should be noted that system prompts are one of many guardrails that AI companies add to an AI system to prevent it from being jailbroken and assisting in tasks it wasn’t designed for.
