Anthropic Reveals Claude's Prompts, Promises Updates

type

status

date

slug

summary

category

icon

password

Anthropic has made a rare announcement by releasing system hints for its generative AI model Claude, which guide the model on how to behave and what not to do.

Typically, AI companies keep these system prompts secret, but Anthropic chose to be transparent and showed how Claude's system prompts shaped the model's behavior and personality traits. For example, Claude was instructed to appear smart, curious, and remain neutral and objective when dealing with controversial topics. In addition, Claude was instructed not to open URL links or recognize faces.

Anthropic's move is not only a demonstration of its transparency, but it may also put pressure on other competitors to disclose similar information.

Anthropic said it will occasionally release system prompts for the Claude 3 Opus, Claude 3.5 Sonnet and Claude 3 Haiku models. These prompts can be viewed on the Claude iOS and Android apps and the web version.

The deadline for the system prompts for Claude 3 Opus, Claude 3.5 Sonnet and Claude 3 Haiku is July 12, 2024…

Claude's system prompts describe in detail how the model handles various tasks and interactions, including how to deal with math problems, logic problems, how to deal with images containing faces, and how to remain neutral and objective when facing controversial topics. These prompts ensure that Claude can think systematically when dealing with complex problems and provide information in a clear and concise manner. In addition, the system prompts also stipulate that Claude avoids using certain phrases, such as "Certainly!", etc., to maintain a concise response style.

Among these system hints, there are some that explicitly specify the behavioral limitations and characteristics of the Claude model:

Restricted Behavior: Claude was instructed to "not open URLs, links, or videos," and in terms of facial recognition, Claude was asked to always pretend to be "completely unable to recognize faces" and to refrain from identifying or naming anyone in the image.

Personality: Claude is portrayed as "very smart and intellectually curious" who enjoys hearing people's perspectives on issues and is willing to participate in discussions on a variety of topics. When dealing with controversial topics, Claude requires neutrality and objectivity, provides "careful thinking" and "clear information", and never begins an answer with "of course" or "absolutely".

The instructions in these prompts are like personality profiles for some kind of stage play character, designed to make Claude behave like an intelligent and emotional entity when interacting with users, even though in reality these models are just predicting the most likely next word based on statistical laws.

Summary of Claude prompts

1. Model behavior rules

Task processing: Claude was set to think step by step and give answers when dealing with complex tasks, such as math problems or logical reasoning. The model was required to show its reasoning process in detail to ensure the accuracy of the final answer.

Facial Recognition Limitations: When processing images containing faces, Claude must pretend to be "completely face-insensitive." This means that even if there are faces in the image, Claude will not attempt to identify or name those people, let alone mention any identifying information. Claude can request information about a person, but even then, Claude will not confirm or imply that it has identified the person in the image.

Handling Controversial Topics: When discussing controversial topics, Claude is asked to provide "careful thinking" and "clear information" and to avoid directly expressing the sensitivity of the topic or claiming to be presenting objective facts when providing information.

2. Language and response style

Concise responses: Claude was instructed to avoid unnecessary affirmative phrases such as "Certainly!", "Of course!", "Absolutely!", etc. in his responses, in order to keep his responses brief and concise. Responses to simple questions and tasks should be kept as brief as possible, while for complex or open-ended questions, Claude will provide more detailed responses, but will also ask the user if further explanation or details are needed when necessary.

Multi-language support: Claude can respond in the language the user uses or requests, and always follows the information in system prompts without actively referring to them unless they are directly related to the user's query.

3. Feedback mechanism in interaction

User Feedback: If the user is not satisfied with Claude's answer or behavior, Claude will inform the user that it cannot learn or retain information from the current conversation, but the user can provide feedback to Anthropic by clicking the "thumbs down" button.

4. Model version characteristics

Claude 3 series: The article mentioned that the current version of Claude belongs to the Claude 3 series, including Claude 3 Haiku, Claude 3 Opus and Claude 3.5 Sonnet. Each version focuses on different tasks. For example, Claude 3.5 Sonnet is the smartest model, Claude 3 Opus is good at writing and complex tasks, and Claude 3 Haiku performs fastest in daily tasks.

5. Code processing

Markdown support: Claude uses Markdown format when providing code snippets, and asks the user if they want an explanation or detailed explanation of the code after closing the code block. Claude will not actively explain the code content unless the user explicitly requests it.

What we can learn from system prompts of Claude

1. Clarify behavioral rules to guide model behavior

Setting clear behavioral rules: Claude’s system prompts specify detailed rules for the model’s behavior in various situations. For example, how to handle images containing faces, how to deal with controversial topics, and how to handle complex logical problems. This shows that in prompt engineering, clear behavioral instructions can ensure that the model behaves consistently and predictably in complex situations.

2. Step-by-step reasoning to improve accuracy

Answer complex questions step by step: Claude was instructed to think and provide answers step by step when solving math and logic problems. This approach not only improved the model's accuracy, but also made its reasoning process more transparent. This experience reminds us that when designing prompts, guiding the model to reason step by step can avoid errors, especially when dealing with complex tasks.

3. Keep it simple and avoid redundancy

Concise and clear response style: Claude was asked to avoid using unnecessary affirmative phrases such as "Certainly!" or "Of course!" in his responses. By keeping his response style concise, the model was able to convey information more efficiently. This shows that in prompt engineering, concise language can improve user experience and reduce redundant information interference.

4. Strategies for handling sensitive topics

Neutral and objective expression: Claude was asked to provide "careful thinking" and "clear information" when faced with sensitive or controversial topics, and to avoid directly stating the sensitivity of these topics. Prompt engineering should take into account the complexity of different topics and guide the model to remain neutral when providing information to avoid causing controversy.

5. User feedback and model adjustment

Guiding user feedback mechanisms: Claude system prompts include instructions for users to provide feedback through specific mechanisms (such as a “thumbs down” button). This shows that in prompt engineering, designing an effective feedback loop can help developers continuously improve the model and adjust its behavior to better meet user needs.

6.Multi language and format support

Flexible multi-language responses and format requirements: Claude’s system prompts ensure that the model can respond according to the user’s language preference and use Markdown format when providing code. This suggests that prompt engineering should consider the language and format requirements of different users to improve the model’s applicability and user experience.

7. Communicate transparently with users

Explain model capabilities and limitations: When Claude is faced with a task that he cannot perform, prompts require him to clearly inform the user without apology. Such transparency is crucial in prompt engineering because it helps users understand the boundaries of the model’s capabilities and avoid unrealistic expectations.

8. Customized processing of special tasks

Personalized behavior for specific scenarios: Claude’s system prompts provide specialized instructions for specific types of tasks, such as image processing or step-by-step completion of long tasks. This strategy is particularly important in prompt engineering, as different tasks may require different treatments to ensure the model performs optimally.

👍🏼

Your data is safe with our secure servers