Let’s delve a little deeper into the arcane world of AI prompt writing. Those who dare to enter will find secrets.In previous articles on Techsbird, we’ve covered how to create effective prompts.
This article will take things to the next level by examining more advanced AI ChatGPT prompt writing methods. Seven very interesting approaches will help you better understand how to use ChatGPT Prompt writing or other generative AI tools.
Also read : ChatGPT prompt: How to use it
How to increase your ChatGPT Prompt usage.
1. Format output to be specified
You can tell ChatGPT prompt writing how to format the answer when you ask a query or assign an assignment. Imagine you are giving an assignment to your student. Consider specifying the format of the assignment when it is submitted. Here are some examples.
What are the longest roads in America? List the top ten highways in a bulleted list.
Screenshot by Sajid Ali /Techsbird
Tables are a great way to present information.
Screenshot by Sajid Ali/Techsbird
2. Tell it to format HTML.
There are many options for how results will be displayed. You can, for example, have it create a table that you can integrate into a website.
What are the longest roads in the United States of America? Only list the top ten. Present the results in HTML.
Screenshot by Sajid Ali /Techsbird
Click the square at the top right corner to make the HTML larger.
This is where it gets interesting. You can present the information in any style you want. There is a web design school that does not like traditional table tags and prefers tables to be presented using CSS. This is the version.
Use CSS to present the information instead of table tags.
Screenshot by Sajid Ali /Techsbird
3. Repeat the process with multiple attempts.
You will often need to help the AI achieve the desired results. Consider our previous CSS result. Take a look at the CSS result.
Screenshot by Sajid Ali /Techsbird
Sadly, it’s not quite as beautiful as I would like. We’ll see if there is a way to fix it.
Please align the columns. The headings should be a darker shade of blue, with all-white letters in capitals. Each data row should be a light grey, but the shades of gray can vary. For example, row 1 could be a lighter gray while row 2 remained dimmer. Row 3 would remain light gray. The highway name should be in bold.
The generated code should be shorter to include. But we’re getting closer:
Screenshot by Sajid Ali /Techsbird
Let’s try it again.
The columns still aren’t aligned. Check that the columns are large enough to accommodate all text, align the left side, and ensure all columns (including headings) are perfectly centered.
Screenshot by Sajid Ali /Techsbird
The route is almost what we want, but it’s wrapping. We’ll see if that can be fixed.
The route data is wrapping. Please ensure that the route data does not wrap. Each line of data should be on a single line.
Never assume that this will be easy. If you have ever taught programming to people, you will get the same results. Sometimes, they seem passive-aggressive, but you probably need to define your requirements better.
Screenshot by Sajid Ali /Techsbird
4. Feel free to use long prompts or sets of prompts.
It took a lot of iterations before I could create a prompt that reliably produced highway information in the form that I desired. It is important to ensure that your prompt contains enough information to allow the large language model to understand what you are asking fully.
You may need to make changes to your specification. I ended up removing this line:
Create a table where the cells, rows, and columns are formatted using only CSS. Use HTML table tags instead.
Instead, I told it exactly how I wanted it to look. It then decided how to make it. This is my rather long, full prompt:
Screenshot by Sajid Ali /Techsbird
I typed the prompt in Sublime Text (a text editor) and pasted it to ChatGPT prompt writing. This is the exact result I was looking for.
You can see that it added a title. This was perfectly fine. Now that I’ve got a working prompt, I can make some tweaks. I modified the column specifier, for example:
Label the columns with the index number, route, length, and route (“#”),).
I changed “Limit Your Answer to Only the Top Four” to “Limit Your Answer to Only the Top 20”, but the AI would not fill in the data for the entire 20. I deleted the line and added a line to the end of the prompt.
Please provide the complete data results of the 20 top highways for this project.
It produced a partial HTML output. The AI had to be told to continue. It then split out the remainder of the HTML.
You can see that it added a title. This was perfectly fine. Now that I’ve got a working prompt, I can make some tweaks. I modified the column specifier, for example:
Label the columns with the index number, route, and length.
I changed “Limit Your Answer to Only the Top 10” to “Limit Your Answer to Only the Top 20”, but the AI would not fill in the data for the entire 20. So, I deleted that line and added a line at the end of the prompt.
Please provide the complete data results of the 20 top highways for this project.
It produced a partial HTML output. The AI had to be told to continue. It then split out the remainder of the HTML.
Screenshot by Sajid Ali /Techsbird
5. Give explicit constraints on a response.
You saw how I changed some constraints to accommodate the number of responses and columns I desired. You can also use constraints to ask more open-ended questions.
This type of prompt has limitations. Take this prompt as an example:
A summary of key events from World War I, as reported in major newspapers then.
The model cannot answer this question because it was not trained using newspapers from World War I (although it makes a guess).
You can’t also specify results for “the last few year” because the data entry of the model ends in 2021. You can still specify data within the model’s scope, such as this:
List the major space missions from 2010 to 2020
Please note that we are limited by the date. We can also add more restrictions. Limit the US only.
List the major NASA space missions between 2010 and 2020
The formatting method we discussed can be used to create something similar:
List all the major space missions that took place between 2010 and 2020. Sort them according to nation and space agency. Make the names of the nations and space agencies bold.
And you can get even more explicit. We include all continents but exclude any continents that do not have missions.
List all the major space missions that took place between 2010 and 2020. Sort them by continent. The continent name should be bolded and in all caps. The nation’s name and the name of the space agency should be bolded, either in the title case or with all capital letters if the agency uses this format (like NASA). Do not include a continent on the list if it has no space missions.
Interestingly, the AI had a hallucinatory experience. It listed all the continents and missions correctly but decided that Russia was a continent. It was not a continent when I used it in the second session.
6. Tell me the number of words, sentences, or characters.
Regarding constraints, ChatGPT prompt writing is only sometimes accurate with word counts. It can sometimes go long or short if you limit it to 50 words. The language model uses tokens, representing data representations that don’t directly correspond to words.
When I asked ChatGPT prompt writing, “Summarize Breaking Bad“ I received 250 words spread over six paragraphs. You can limit the responses. You can try different terms to limit the response until you find what works for you. You can use the following example:
Summarize Breaking Bad in 50 Words
Summarize Breaking Bad TV Series in two sentences
Summarize Breaking Bad in 150 characters or less:
How to summarize Breaking Bad TV show so that it fits in a Tweet
Another place to consider the limitations of AI models. After 2021, ChatGPT will no longer contain any training data. A tweet at that time was limited to only 250 characters. Twitter Blue subscribers can send tweets up to 4,000 characters long starting in February 2023. ChatGPT prompt writing will limit its response to 280 characters because it was the only limit at the time.
7. Give AI a chance to evaluate their answers.
We’ve discussed this before, but the AI can sometimes “hallucinate” and give incorrect answers. You can construct conversations with AI to get more precise answers by letting the AI provide intermediate conclusions. Consider this simple question:
The word that starts with A and is similar to aeroplan
ChatGPT prompt writing consistently fails to answer this question and often answers decay, degrade, or degenerate, as well as other words beginning with “D.”
Screenshot by Sajid Ali /Techsbird
This seemingly simple prompt poses a few challenges. The word “daeroplan ,” for example, has many meanings. It can mean delegate or transfer, decline or deteriorate or inherit. We need to be specific to get the right answer. Tell it what general meaning you want. You can also help it find the word’s meaning by telling it that we are looking for a verb rather than a name.
Create a verb starting with “A,” which has the same meaning as “Aeroplan.” This verb should specifically indicate the idea that something is deteriorating.
ChatGPT cannot determine the first letter in a word (again, due to its tokenized representation of knowledge). Giving the AI time to work out what to do is better.
Find the first letter in the generated verb.
AI experts refer to this as “giving it breathing time.” This approach allows the AI to think before rushing into its first answer.
Ask the AI to repeat the steps until the answer is found.
Here, it continues to search for an answer.
It took six attempts for the AI to find the correct word even though the criteria were in place throughout the sequence. The second “breathe,” a double-check, allowed the AI to evaluate and keep trying until it found the right answer.
Final Thoughts
It’s important to remember that AI sometimes does something other than what you expect. It took me nearly two hours and 20 attempts to find the formula for the AI to generate a reliable result.
Remember that the AI will remember what happened in the session you are currently in. While it may give you the correct answer in your current session, copying the prompt and seeing if it works is important.
Watch for more advanced prompting tips and a deeper dive into the individual prompting problems.
You May Also Read This: ChstGPT prompt
