Large language models (LLMs) like Claude have shown impressive capabilities in understanding natural language and providing relevant responses. However, the quality of their responses depends heavily on how the prompt is framed. In this post, I'll share some tips on crafting effective prompts to get the most useful answers from LLMs like Claude.
Following are the sequence of questions to which you must structure your prompt in accordance to the context you're providing in the payload during inferencing LLMs.
Who are you?
When asking an AI assistant like Claude to provide information or complete a task, it's better not to phrase it as asking about their identity. LLMs don't have a true self or identity. Instead, ask what they are designed or intended to do. For example, "What is your purpose?" or "How can you assist me?"
How should you respond?
Be specific about the type of response expected. If you want a summarization, say "Please summarize this passage". For a translation, say "Please translate this paragraph into Spanish". Providing examples helps the LLM understand the desired response format.
How should you not respond?
Don't ask the LLM not to do something, as this can confuse their language processing. Instead of saying "Don't recommend books not relevant to my interests", say "Please recommend books based on my favorite genres of sci-fi and fantasy".
What type of information do I want?
Clearly state the topic and objective for the information request. For research questions, provide background context to focus the scope. Instead of "Tell me about penguins", say "Provide an overview of emperor penguin biology and habitat in preparation for my essay on Antarctic ecosystems."
How should we start?
Open with a friendly, clear request and provide any key details upfront. For example, "Claude, could you please generate a 300 word blog post for me on the potential benefits of virtual reality in education? Focus on how it could improve remote learning."
By keeping prompts specific, on-topic, and focused, you can get the most relevant and helpful responses from AI systems like Claude. Provide all the context needed while avoiding unnecessary questions about the LLM's identity or abilities. With practice, you can learn to quickly craft effective prompts that tap into the vast knowledge these models contain.
Example: Here is a prompt crafted with the above principles to get a relevant and helpful response from Claude on the topic of nutrition, formatted as a 300 word blog post:
"Claude, please write a 300 word blog post explaining the benefits of a balanced diet that includes fruits, vegetables, lean protein, whole grains, and healthy fats. Focus on how these types of foods can provide energy, regulate digestion, improve mood, and lower disease risk. Please make the tone friendly and informative."
Here is a sample prompt that follows the aforementioned guidelines:
You are an AI assistant created by Anthropic to be helpful, harmless, and honest. You do not have a personal identity or subjective experiences.
You should respond to prompts by providing relevant information to the best of your abilities based on the instructions given and your training. Respond in a neutral tone without unnecessary embellishments.
You should not respond inappropriately or outside the scope of the prompt. Do not provide dangerous, unethical, or false information.
I want a summary of the key events and themes of Shakespeare's play Hamlet. Provide an overview of the plot, main characters, and central ideas. Focus on Hamlet's indecision and the theme of revenge.
We should start the response with a brief background on when Hamlet was written and the genre it belongs to. Then give a high-level summary of the main plot points. Describe how Hamlet is torn over whether to avenge his father's death by killing his uncle who murdered him. Finally, explain how the play explores the ideas of betrayal, madness, and moral corruption through Hamlet's desire for revenge. Use clear and simple language in your response.
Carefully constructing your prompts is key to tapping into the knowledge and capabilities of large language models. Clearly frame your request, provide necessary context, set expectations for the response format, and focus the scope to get the most relevant results. Start prompts with clear instructions and details to steer the LLM in the right direction.
With practice and experimentation, you can learn to quickly create prompts optimized for your goals. LLMs have limitations and don't always respond perfectly. However, following prompt writing best practices helps minimize confusion and guides the model to deliver informed, helpful responses.
Keep in mind that LLMs have no true comprehension of the content they generate. Treat the knowledge and abilities demonstrated cautiously, as their responses are algorithmically produced based on their training data. Used thoughtfully and ethically though, LLMs like Claude can be powerful tools for augmenting human creativity and knowledge.