From Meh to Marvelous: The Secret Sauce of Stellar AI Prompts

Prompts can make or break a successful interaction with an AI.?The importance of well-crafted prompts cannot be overstated. They can differentiate between vague, irrelevant responses and precise, insightful answers.?

Since the release of our custom GPT?Openbridge Data Pilot, we are often asked?how to craft effective prompts.

Prompt authoring is not a one-size-fits-all discipline. There are foundational elements for good prompts that can enhance information retrieval accuracy, improve the creativity of generated content, and even help navigate complex problem-solving scenarios.

We will cover several prompt writing patterns, including RAIL and PTCF for Open AI and Google.

RAIL and PTCF: Two Approaches to Prompt Engineering

There are various prompt writing approaches and frameworks, each with strengths and weaknesses. Two common prompt patterns are RAIL (Retrieval-Augmented Instruction Learning) and PTCF (Persona, Task, Context, Format). OpenAI often suggests RAIL, while Google Gemini usually references PTCF.

RAIL is centered around providing relevant reference information to the AI, enhancing its ability to draw upon specific knowledge. PTCF, on the other hand, is a structural framework that helps craft more comprehensive and targeted prompts.

While these approaches differ, they aim to improve AI outputs.

Let's dive in.

Understanding PTCF: Persona, Task, Context, Format

PTCF is a framework for structuring prompts, considering four key elements: Persona, Task, Context, and Format. This approach helps create more focused and effective prompts.

Pros

• Provides a clear structure for crafting prompts
• Improves consistency and relevance of responses
• Allows for fine-tuned control over the AI's output

Cons

• Can make prompts longer and potentially more complex
• It might require more time to craft initially
• It may limit creativity if applied too rigidly

Example:

## Persona:
- As an E-commerce data analyst, you are tasked with evaluating advertising campaign performance for a mid-sized online retailer.

## Task:
- Analyze the effectiveness of various Amazon advertising campaigns using data from the "amzn_stream_campaigns" table accessed through the Openbridge API.

## Context:
- The analysis will focus on campaigns involving Sponsored Products and Sponsored Brands over a specific time period. 
- The goal is to identify high-performing campaigns and areas of underperformance, considering external factors that might influence results.

## Format:
Results from the analysis will be summarized in a report that includes:
- Assessments of sales versus advertising spend.
- Trends in click-through rates (CTR) across different campaign types.
- Seasonal or event-driven impacts on campaign performance.
- Metrics such as Total Advertising Cost of Sales (ACoS), CTR, Conversion Rate, and Return on Advertising Spend (ROAS).
- Actionable recommendations for budget reallocation and strategic adjustments based on observed ROAS and CTR outcomes.        

This prompt is a starting point to help frame the context of the conversation. Like any exploratory conversation, iteration is critical.

Iteration Strategies

1. Refine the persona: If the tone or expertise level isn't right, adjust the persona description. For example, change from "senior software developer" to "software architect focusing on scalability."
2. Clarify the Task: If the instructions are too broad or lack specificity, ask for more detailed guidelines or break down the task into smaller, more focused subtasks. For example, instead of "Analyze sales data," specify "Analyze the monthly sales trends for the top 5 products in the US market using the sales_data_master table."
3. Ensure Data Accuracy: Always call the Openbridge API to retrieve the latest dataset for the sp_orders_report_master table to ensure you are working with the most up-to-date information.
4. Expand or narrow the context: If the response doesn't match the intended audience's needs, provide more background information or specify the audience's knowledge level more precisely.
5. Modify the format: If the response's structure isn't ideal, adjust the format instructions. For example, change from paragraph form to bullet points or specify a different number of sections.
6. Use examples: If the model's output style isn't quite right, provide a short example of the desired output format or style.
7. Iterative prompting: Use the output from one prompt as input for a follow-up prompt to refine or expand on specific points.

Exploring RAIL: Retrieval-Augmented Instruction Learning

RAIL is a technique that provides relevant reference text to the model to improve the accuracy and reliability of its responses. This method leverages external knowledge to supplement the model's training data.

Pros

• Improves accuracy by providing up-to-date or specific information
• Reduces hallucinations or made-up information
• Allows for domain-specific knowledge injection

Cons

• Requires a reliable source of relevant information
• Can increase prompt length and potential costs depending on the context
• May slow down response time due to additional processing

Example:

## Results:

- Begin by accessing the amzn_stream_campaigns table from the Amazon Marketing Stream - Campaigns through the Openbridge API.
- Evaluate the ROI by analyzing which campaigns have generated significant sales relative to their advertising spend.

## Actions:

- Identify underperforming campaigns and suggest reallocation of budgets to those with higher ROAS.
- Recommend experimenting with new ad formats or targeting strategies for sponsored products with lower CTR.

## Insights:

- Examine trends in CTR across Sponsored Products and Sponsored Brands.
- Assess the impact of seasonal factors or external events on campaign performance.

## Learnings:

- Determine critical metrics like Total Advertising Cost of Sales (ACoS), Click-through Rate (CTR), Conversion Rate, and Return on Advertising Spend (ROAS).        

Again, this is a starting point for a conversation, so iteration and experimentation are key.

RAIL IIteration Strategies

1. Refine the reference text: If the model isn't providing accurate answers, try selecting more relevant or comprehensive reference material.
2. Adjust specificity: If answers are too general, instruct the model to be more specific. Conversely, if answers are too detailed, ask for higher-level summaries.
3. Use follow-up queries: If the initial response is incomplete, ask follow-up questions to gather more information.
4. Modify the instruction: Experiment with different phrasings or levels of detail in your instructions to guide the model more effectively.

Common Pitfalls: Examples of Ineffective Prompts

Clear communication with AI is a must. Whether using RAIL or PTC, the focus should always be guiding the AI toward producing the most valuable and accurate responses possible.

Here are a few "bad" examples of prompts that lack clarity or purpose, making it difficult for the AI to respond with meaningful outcomes.

RAIL (Retrieval-Augmented Instruction Learning) Bad Examples

• Vague and Unfocused

Look at the Amazon data and tell me what's going on with the campaigns.        

This prompt needs more specific instructions and provides reference information, defeating the purpose of RAIL.

• Detailed Information With No Clear Outcome

Analyze the amzn_stream_campaigns table from the Amazon Marketing Stream - Campaigns through the Openbridge API. Look at ROI, sales, advertising spend, CTR, ACoS, ROAS, conversion rates, seasonal factors, external events, ad formats, targeting strategies, budget allocation, and everything else you can find. Give me a complete breakdown of all this information.        

This prompt provides too much information without clear prioritization, potentially overwhelming the AI and leading to unfocused results.

• Lack of Specific Actions

Check the Amazon sales data tell me how things are going        

This prompt needs to provide specific actions or insights derived from the data, making it difficult for the AI to generate functional responses.

PTCF (Persona, Task, Context, Format) Bad Examples

• Incomplete PTCF Structure

Persona: You're an analyst.

Task: Look at some Amazon data.

Context: It's about sales advertising.

Format: Tell me how my business is doing.        

This prompt uses the PTCF structure but needs meaningful information in each section, resulting in vague and unhelpful guidance.

• Mismatched Persona and Task

Persona: As a junior intern with no experience in data analysis,

Task: Provide an in-depth analysis of complex Amazon advertising campaigns using advanced statistical methods.

Context: The company needs this analysis for a board meeting tomorrow.

Format: Create a 50-page report with detailed charts and projections.        

This prompt creates an unrealistic scenario by assigning a complex task to an inexperienced persona, likely leading to poor-quality output.

• Lack of Specific Context and Format

Persona: You are a marketing expert.

Task: Analyze some Amazon campaigns.

Context: It's for an online store.

Format: Make it look professional.        

This prompt needs more specific context about the campaigns and needs to provide clear formatting instructions, potentially resulting in generic and unhelpful responses.

These examples demonstrate how poorly constructed prompts can lead to vague, unfocused, or inappropriate responses from AI models.

My Prompts Always Suck, Now What?

Sometimes, writing good prompts can be challenging; it can be a process of trial and error, even for pros. However, you can use AI to help you construct "good" prompts.

The beauty of prompt engineering lies in its flexibility – you can experiment with these methods, combine them, or even develop variations to suit your specific needs.

Dr. Lance Eliot , an expert on artificial intelligence (AI) and machine learning, wrote a great guide on honing your prompt writing skills with assistance from the AI you are entering prompts into! The article covers strategies and techniques for building a core set of approaches to creating, integrating, and employing prompts more effectively.

• The Clever ‘Rephrase And Respond’ Prompting Strategy Provides Big Payoffs For Prompt Engineering

Mastering the Art of Prompt Crafting: Key Takeaways

Formulating your prompt takes time, and regardless of the specific methodology employed, it maximizes the potential for productive AI interactions.

Whether using RAIL or PTCF, the ultimate goal is to craft prompts that elicit precise, relevant, and insightful responses from AI models. The key to success lies not in rigidly adhering to one approach but in understanding the principles behind effective prompts and flexibly applying them to suit specific needs.

We published some thought starter prompts in our Openbridge Data Copilot project documentation. These are not "cut-and-paste" prompts but are meant to provide starting points to begin your journey.

Don't hesitate to try multiple variations and combine different elements of RAIL and PTCF to achieve the best results. This approach leads to better outcomes and accelerates your time to value.?

References

• Google Gemini Prompting Guide 101: A quick-start Handbook for effective prompts
• OpenAI Prompt Engineering
• Openbridge Amazon Data Copilot