Unlocking the Potential of ChatGPT: A Powerful Assistant for Content Creation Part 2

Be Specific and Structured with Your Prompts

ChatGPT does its best when you create and use specific, logically structured prompts. I’ve also discovered that I get the best results when I break complex prompts into smaller, discrete tasks. There are techniques for using much larger prompts, but I prefer my approach for two reasons:

1. I can see the results one step at a time and adjust the current prompt or the following ones.
2. Complex prompts can generate lots of data you must wade through to find what you need or want.

My approach involves writing the prompt from start to finish and then breaking it into individual tasks. Here are some things to keep in mind when you write your prompts:

• Tell ChatGPT to use both its knowledge base and the internet to generate a response. The ChatGPT knowledge base is several months old, meaning it won’t provide the most current results if you don’t ask for them. Example:

Hi Chat, Please use your knowledge base and the Internet to find the best technologies for precision machining composite materials.

• Tell ChatGPT to provide specific responses. Otherwise, it will return generalized results in most cases. Example:

Hi Chat, I need to develop a list of best practices for mining cryptocurrencies. Using your knowledge base and the Internet, please provide specific examples of the current best practices.

• Ask for the URL of the references used in the response. This helps verify accuracy. If you forget, do not worry; ask for the URLs in a follow-up prompt. Example:

Thank you for the best practices for cryptocurrency mining. Could you please provide me with the URL for each best practice?

• You can ask ChatGPT to return the results in a table. I find this particularly useful in my writing. Example:

I am interested in comparing the lifting capacity of the English, African, and North American Sparrow. Specifically, I want to know how much weight each can carry and for what distance. Please put the results in a table.

• Ask ChatGPT if it understands your prompt and what you want in the results. Example:

I am interested in comparing the lifting capacity of the English, African, and North American Sparrow. Specifically, I want to know how much weight each can carry and for what distance. Please put the results in a table. Before you begin, do you understand my prompt? If not, please tell me what else you need.

• ChatGPT can only access so much data at a time. So, if your chat session is extensive, ChatGPT will forget earlier content to have enough room to address current requests. The solution is simple: ask ChatGPT to reference responses or input provided earlier. Example:

Hi Chat. I need to limit my blog posts to 3000 characters. Could you please analyze my Part 1 copy provided earlier in this chat and break it down into 2500-character chunks? That gives me room for transitions:>)

• Ask ChatGPT to help you write the prompt to allow it to deliver the best results. Example:

Hi Chat, I’d like you to help me create the perfect prompt so that you can give me the results I need. Please provide the prompt you’d like to see if I am seeking data about the weight-carrying capabilities and flying distances with the load of several types of sparrows.

• ChatGPT can’t count, at least not very well. So, if you are looking for an exact number of references or words in response to your prompt, you aren’t likely to get it. Strangely enough, ChatGPT can be pretty good at math. Go figure.
• ChatGPT doesn’t always spell correctly when you ask for help generating an image or graphic using DALL-E. However, it is improving, as the featured image for this part shows.

Use these techniques, and you’ll get great results from ChatGPT every time. Don’t be shy about reworking a prompt if you do not get what you want the first time, the second, or…oops! ChatGPT might be hallucinating!

Hallucinations in ChatGPT Results

You may have read about the issue of LLMs hallucinating and producing wildly out-there results. No, these aren’t chemically or trance-induced; they happen when prompts are incomplete, vague, or otherwise incorrectly written. I thought it would be an excellent exercise to ask ChatGPT to help craft this section of Part 2. Let’s start with the definition of an LLM hallucination:

USER:

Hi Chat. Happy Hump Day! I am working on the second part of my LinkedIn using the ChatGPT series. I'd like a concise, non-jargon definition of LLM hallucinations if you please.

ASSISTANT:

Happy Hump Day! Here’s a concise, non-jargon definition of LLM hallucinations for you:

LLM Hallucinations are instances where large language models, like ChatGPT, generate information or responses that are inaccurate, misleading, or completely fabricated, even though they sound plausible. This happens because the model predicts text based on patterns in the data it was trained on rather than verifying facts.

Let me know if you need any adjustments!

Then, I asked ChatGPT for a famous example of a hallucination. Here’s what I received.

USER:

Cool, thanks! It would be fun to include a hallucination as an example. Do you have a favorite or know of a famous example? If this is insensitive, please let me know, and I won't ask you to do it.

ASSISTANT:

It's not insensitive at all! Using examples can be a great way to illustrate the concept. Here's a well-known example of an LLM hallucination:

Famous Example of LLM Hallucination

One of the most famous examples occurred with OpenAI's GPT-3. When asked about the history of the fax machine, GPT-3 generated a detailed and plausible-sounding response that included the claim that the fax machine was invented by Alexander Graham Bell, who is known for inventing the telephone. In reality, the fax machine was developed by Alexander Bain in the 1840s.

Example Hallucination:

User: Who invented the fax machine?

GPT-3 (Hallucination): The fax machine was invented by Alexander Graham Bell, who is also known for inventing the telephone. Bell's invention of the fax machine revolutionized communication by allowing documents to be sent over telephone lines.

This example shows how the model can generate incorrect information that seems credible, illustrating the importance of verifying facts and providing clear prompts to minimize hallucinations.

Finally, I asked ChatGPT to verify the two primary techniques I use to combat hallucinations and then asked for additional techniques it would recommend.

USER:

I understand that the best way to avoid hallucinations is to ask you for specific results and to ask you if my prompt is clear and understandable. Do these techniques help?

ASSISTANT:

Yes, those techniques do help. Here’s how they work:

Specific Results

• Clarity: Asking for specific results ensures that the model focuses on a narrow, well-defined topic, reducing the chances of generating irrelevant or inaccurate information.
• Precision: Specific prompts guide the model to provide detailed and accurate responses, which can help in minimizing hallucinations.

Checking Prompt Clarity

• Feedback Loop: By asking if your prompt is clear and understandable, you create a feedback loop where the model can confirm whether it has correctly understood your request.
• Refinement: This allows you to refine your prompts for better accuracy, ensuring that the model’s output aligns more closely with your expectations.

Using these techniques can significantly reduce the occurrence of hallucinations and improve the quality of the responses you receive.

USER:

Thanks! Are there other techniques you recommend for preventing hallucinations?

ASSISTANT:

Certainly! Here are some additional techniques to help prevent hallucinations in large language models (LLMs):

Techniques to Prevent LLM Hallucinations

1.      Provide Context:

• Detailed Background: Give as much relevant information as possible in your prompt to provide context. This helps the model generate more accurate responses.
• Examples: Use specific examples to clarify what you are asking for.

2.      Use Multiple Queries:

• Iterative Approach: Break down complex questions into smaller, more manageable parts and ask them sequentially.
• Cross-Verification: Ask the same question in different ways to compare the responses for consistency.

3.      Ask for Sources:

• Cite References: Request the model to provide sources or references for the information it generates.
• Verification: Check the provided sources independently to verify the accuracy of the information.

4.      Set Clear Boundaries:

•  Define Limits: Explicitly state any constraints or boundaries for the response, such as a specific time period, geographic location, or type of information.
• Reinforce Parameters: Reiterate key parameters within the prompt to ensure the model stays focused.

5.      Prompt Repetition:

• Repeat Key Points: Emphasize important aspects of your request multiple times within the prompt to reinforce understanding.
• Summarize Requests: Summarize the main points at the end of the prompt to ensure clarity.

6.      Seek Clarification:

• Interactive Dialogue: Engage in a back-and-forth conversation to refine and clarify the information.
• Ask for Elaboration: Request the model to explain its reasoning or provide more detail if an answer seems unclear.

7.      Use External Verification Tools:

• Cross-Check: Utilize external tools and resources to cross-check the information provided by the model.
• Fact-Checking Services: Leverage online fact-checking services to validate critical information.

By incorporating these techniques, you can enhance the accuracy and reliability of the responses you receive from LLMs like ChatGPT.

Try It Yourself

One of the cool things about learning to use ChatGPT and other AI tools is how easy it is to get started. If you want to test any of the queries in this article, all you need do is copy the content following USER and paste it into the chat field as shown below.

Or you can use the handy template I asked ChatGPT to create to write excellent, general-purpose prompts. After ChatGPT created the first template, I realized I hadn’t asked for one that would avoid hallucinations and would deliver specific, current results. You can see my modified request and the resulting template below.

USER:

Thanks! Before I go, what changes would you make to your template to help you avoid hallucinations and provide specific, current results?

ASSISTANT:

To help avoid hallucinations and provide specific, current results, you can make a few tweaks to the template:

1. Ask for Verification of Sources: Request the AI to mention its sources or specify where the information is derived from.
2. Specify Date for Current Trends: Request information on recent trends within a specific time frame.
3. Encourage Clarity in Prompts: Ask the AI to confirm if it needs more details or if the prompt is clear.

Here's the updated template:

Enhanced General-Purpose Research Prompt Template

Prompt:

Hi ChatGPT, I hope you're doing well. I need your help with some research on [Topic]. Could you please provide me with the following information:

1. Overview: A brief summary of the topic.
2. Key Points: Important aspects or subtopics related to the main topic.
3. Current Trends: Any recent developments or trends in this area, specifically from the past [time frame, e.g., 6 months].
4. Sources: Reliable sources or references where I can find more detailed information. Please specify the sources or references for the information provided.

Please confirm if this prompt is clear, and let me know if you need any additional details to provide accurate and up-to-date information. Thank you!

Example Usage:

Prompt:

Hi ChatGPT, I hope you're doing well. I need your help with some research on renewable energy. Could you please provide me with the following information:

1. Overview: A brief summary of renewable energy.
2. Key Points: Important aspects or subtopics related to renewable energy.
3. Current Trends: Any recent developments or trends in this area, specifically from the past 6 months.
4. Sources: Reliable sources or references where I can find more detailed information. Please specify the sources or references for the information provided.

Please confirm if this prompt is clear, and let me know if you need any additional details to provide accurate and up-to-date information. Thank you!

These additions encourage the AI to focus on current and verifiable information, enhancing the accuracy and relevance of the responses.

Time to Test the Water

First things first, you’ll need a ChatGPT account or one with your LLM of choice. ChatGPT offers free and paid options. Since I am a heavy user, I subscribe to ChatGPT, which enables me to use its latest and greatest models. Check it out at https://openai.com/. If you want to test the waters, set up a free account. Open AI explains the pros and cons pretty well.

Sign into your account, and in the upper lefthand corner, select the model you want to use (#1), then click the new chat button (#2).

You are now ready to engage with ChatGPT! The prompt field (#3) is at the very bottom of your screen. Click in the field, type your prompt, hit Enter, and ChatGPT starts working. Here’s a hint that works for me: I type out my prompt in Word and copy and paste it into the prompt field.

Provide Feedback

Providing feedback during interactions with ChatGPT and other LLMs is important. While the feedback doesn’t directly impact new models' ongoing training and development, developers use the input and common questions to guide their development efforts. You provide feedback in two ways.

1. I reply with positive feedback by typing directly into the prompt field. My responses are short, like “Thank you,” “Perfect, just what I needed,” and so forth.
2. For negative feedback, I click on the little thumbs-down icon shown above #4. A small dialog box asks for feedback.

Conclusion

Regarding prompts, the old programmer saying, “garbage in, garbage out,” holds true. To get the most out of ChatGPT acting as your content creation assistant, you must master the art of writing prompts. In certain circles, this is known as prompt engineering. Here are my suggestions for writing great prompts:

• Clearly define what you want and ask for specific results.
• Don’t ask for too much in a single prompt.
• Ask if ChatGPT understands the prompt.
• Ask for source URLs.
• Be pleasant in your interactions.
• Have fun!

Stay tuned for Part 3, where I'll delve into how I use ChatGPT to augment and improve my content without letting it do the writing.

Disclaimer: The content of this article series, including all text and images, is protected by copyright law. Unauthorized use or duplication of this material without express and written permission from the author is strictly prohibited. For permissions, contact me at wccmn@outlook.com. Hey, I may want to make it into a movie! Thanks.