Prompts Strategies

Begin with the Basics

Embarking on prompt design is inherently a process of trial and refinement, demanding ample experimentation to pinpoint the most effective strategies. Starting with straightforward prompts in a development environment like OpenAI's or Cohere's playground can serve as a solid foundation.

Initiate with basic prompts and progressively incorporate additional elements and context to enhance outcomes. The iterative refinement of your prompts is crucial. Throughout this guide, you'll discover that specificity, simplicity, and brevity often lead to superior results.

For complex tasks comprising multiple subtasks, consider deconstructing them into simpler, more manageable components. This strategy helps prevent overwhelming complexity early in the prompt design process.

‍

Crafting the Prompt

For a range of straightforward tasks, commands can direct the model towards your intended goal, employing verbs like "Write", "Classify", "Summarize", "Translate", "Order", etc.

Experimentation is key. Test various instructions, contexts, and data to ascertain the most effective approach for your specific scenario. Generally, the more tailored and pertinent the context is to your task, the better the outcomes. We'll delve into the significance of context and experimentation in subsequent guides.

A common practice is to position instructions at the prompt's outset, potentially using a clear separator like "###" to delineate instruction from context.

‍

Example:

Prompt:

Instruction:
Translate the following text into Spanish: Text: "Good morning!"

‍

Output:

¡Buenos días!

‍

Emphasize Specificity

The clarity and detail of your instruction are paramount. A descriptive and detailed prompt typically yields better results, especially when aiming for a specific output style or format. While there's no magic formula of tokens or keywords, a well-structured and descriptive prompt is invaluable. Including examples within your prompt can significantly enhance the specificity and relevance of the output.

However, be mindful of the prompt's length due to inherent limitations. Overloading your prompt with unnecessary details can be counterproductive. The included details should directly contribute to achieving the task, necessitating considerable experimentation to find the optimal balance.

‍

Example:

Prompt:

Identify and list the names of places mentioned in the text below.Desired format:
Places: <comma_separated_list_of_places>
Text: "Despite the challenges, the project's success in the Silicon Valley has been noteworthy. 
John Doe, a software engineer at Tech Innovations Inc. in San Francisco, commented on the breakthrough."

‍

Output:

Places: Silicon Valley, San Francisco

‍

Avoid Vagueness

While striving for detail and improved formatting, avoid overly clever or vague prompts. Direct and specific prompts tend to communicate your intent more effectively, akin to clear and effective communication.

For instance, if exploring the concept of prompt engineering, a vague prompt might be:

Explain prompt engineering briefly.

This lacks clarity on length and style. A more effective prompt would be direct and concise:

In 2-3 sentences, describe prompt engineering to a novice programmer.

‍

Directness Over Negation

When designing prompts, focus on what the model should do rather than what it shouldn't. This promotes clarity and specificity, leading to more accurate responses from the model.

‍

Example of a less effective prompt:

Prompt:

This is a bot that suggests movies. 
DO NOT inquire about the user's preferences or personal details.
User: Suggest a movie that aligns with my interests.Bot:

‍

Output:

Certainly! To tailor my recommendation, could you share your favorite movie genre?

‍

Example of a more effective prompt:

Prompt:

This bot recommends movies based on current global trends, 
without soliciting user preferences or personal information. 
If unable to recommend, it should say, "Sorry, no recommendations available today."

User: I'm looking for a movie suggestion.
Bot:

‍

Output:

Currently, I don't have specifics on your preferences. 
However, here are some globally trending movies: [movie list]. 
Enjoy exploring these options!

‍