Introduction

During my work on writing prompts for the Earthmera eco-action detection LLM model, I discovered that even small changes in the prompt could significantly impact the model's performance. Initially, I wrote the prompts without any specific guidelines or structure, relying solely on plain English sentences. However, the results were disappointing and fell short of my expectations. This led me to delve deeper into the field of prompt engineering, where I learned various techniques to enhance the model's performance. In this blog post, I'll share some of the key strategies I discovered and applied to optimize prompts effectively.

Different types of prompt techniques can be used depending on the use case. LLM models can be employed for various tasks, such as Question & Answering, Text summerization, Text classification, Code Generation, and more. Additionally, with multi-modal LLM models, it's possible to handle tasks like image/video generation, summerization, classification, and others. Therefore, it’s essential to understand which techniques to apply for each task to get the most accurate and relevant responses from the model.

Guidelines

There are several guidelines you can follow to achieve better results:

Start simple
Use concise expressions
Provide clear instructions
Specify task conditions
Provide context
Specify the output format

Start simple

Begin with a basic prompt and gradually refine it by adding necessary elements and removing anything unnecessary. This approach helps you focus on the core task and makes it easier to identify what improves or detracts from the model’s performance.

Initial prompt

Explain climate change.

Refined prompt

In simple terms, explain the causes and effects of climate change,
including its impact on ecosystems and human societies.

Use concise expressions

When writing prompts, be as brief as possible without losing meaning. Long, complex prompts can confuse the model, while concise ones ensure it focuses on the core task.

Verbose prompt

Can you please provide an explanation of the greenhouse effect and
how it contributes to global warming, detailing the gases involved?

Concise prompt

Explain the greenhouse effect and its role in global warming,
including the gases involved.

Provide clear instructions

Ensure that your prompt clearly states what you want the model to do. Ambiguous instructions can lead to poor results or incorrect outputs.

Unclear prompt

Tell me about pollution.

Clear prompt

Describe the main sources of air pollution and their impact on human health.

Specify task conditions

If there are specific requirements or constraints for the task, make sure to explicitly mention them in the prompt.

Unspecified conditions

Summarize the article.

Specified conditions

Summarize the article in 50 words, focusing on the main findings and their implications.

Provide context

Giving the model some context helps it understand the situation better and generate more accurate responses.

Without context

List ways to save energy.

With context

List ways to save energy in a household setting to reduce electricity bills.

Specify the output format

To ensure the output is useful, specify the format in which you want the answer. This is especially helpful for tasks like data generation or structured outputs.

Without format

List the countries affected by deforestation.

With format

List five countries affected by deforestation, in bullet points.

Techniques

By following the guidelines above, the model will provide good-quality answers. However, to achieve the best quality responses, you may need to apply additional techniques. I used following techniques when I write the prompt for Earthmera eco-action detection model:

Generated knowledge prompting
One-shot prompting

Generated knowledge prompting

This technique involves providing relevant knowledge or information along with the question to help the model produce more accurate responses. By using this method, it’s possible to improve the model's reasoning abilities while maintaining its flexibility.

Generated knowledge prompting is used in the prompts for the Earthmera eco-action detection model to provide more context on what the model should look for in the input image or video, helping it make more accurate determinations.

I can't share detailed information about the specific prompts we use, but here’s an example of how we apply the generated knowledge prompting technique. If we want the model to determine whether a user is using an eco-product in a video, we first need to define what an eco-product is for the model.

An eco-product is an environmentally friendly item designed to minimize negative impacts on the environment.
These products often use sustainable materials, have minimal packaging, or promote energy efficiency.
Examples include reusable water bottles, biodegradable utensils, solar-powered devices, or items made from
recycled materials. In the video, identify if the user is interacting with or using any of these eco-products.

I apply this generated knowledge prompting technique whenever the model needs a definition or specification for certain materials or products. It helps the model understand what to focus on in the image or video.

One-shot prompting

One-shot prompting is a technique where you provide the model with a single example to guide its response. This helps the model understand the task by showing it a relevant instance, making it more likely to generate accurate results.

In my use case, I use one-shot prompting to provide an example of the output format, ensuring that the model’s response can be used in the subsequent pipeline of the service.

...
Provide your answer in the format: YES, [carbon reduction in grams] or NO.
Example 1: YES, 200
Example 2: NO

In addition, techniques like prompt chaining, Retrieval Augmented Generation (RAG), and Chain-of-Thought (CoT) prompting can be used to further improve the accuracy of the model’s responses. I plan to study more about these techniques through research papers and post about them on the blog in the future.

Prompt engineering techniques