Introduction
On November 3, 2022, GPT-3.5, a tool that revolutionized the way we think about content creation and launched generative AI as a disruptive trend for professionals, was made public, becoming a cornerstone of modern productivity.
Even after a year, many still struggle on unlocking its full potential, either due to skepticism or a simple lack of understanding.
To exploit it properly we need to master specific techniques, called prompt engineering, and some kind of talent in the ability to analyze, transform and extract the information the machine can provide us.
As someone who has basically replaced Google with ChatGPT in my daily routine, I feel compelled to share my insights to celebrate its anniversary.
Disclaimer
My goal is to equip anyone, from absolute beginners to those seeking more advanced knowledge, with the tools to be proficient with ChatGPT.
The guide is designed to be gradually more complex, trying to bring everyone closer to my daily usage of this machine.
Because of that there are many simplifications and omissions, to keep this resource as user-friendly as possible.
I want to make the power of AI accessible to a broad audience without overwhelming the newcomers.
All the insights inside this guide are valid for the free version of ChatGPT.
That said, I highly recommend investing in the premium account at $20/month.
In my professional day-to-day, I utilize many paid tools, but the investment in the premium features is by far the most rewarding.
Please keep in mind that all the information is accurate as of November 2023.
While I may update this guide or create additional content in the future, is crucial to stay informed about the fast-evolving landscape of AI.
Navigation menu
What’s ChatGPT and its utility
There is a fundamental premise: ChatGPT is not "intelligent". It does not possess any reasoning capabilities.
It's a Large Language Model (LLM), trained on vast quantity of text, designed to generate the most probable response to our questions.
Given its probabilistic nature, GPT is not infallible, it can produce errors or trivial responses.
Because is trained on an extensive, but still finite dataset, GPT is not born to be considered as an all-knowing oracle.
GPT will always try to give you a response, but without reasoning, the art of crafting the right question is crucial to get the right answer.
Since its release, ChatGPT has been utilized for a myriad of purposes:
it’s a writing assistant, an idea validator and generator, a brainstorming tool, a teacher, an instrument for summarizing content and crafting documents.
This guide will equip you with the skills to utilize ChatGPT in these roles, with applications that span potentially across all the professional fields.
The premium version of ChatGPT currently allows access to the model GPT-4 and internet browsing, image analysis and creation, chart crafting, and the use of external plugins, but that will not be part of this guide.
More complex uses, like programming via ChatGPT or automating business processes, are for more advanced users and would require dedicated articles.
Let's unlock the full potential of this tool together and become a pro on using ChatGPT!
Crafting Your First Command
The first step is to access to the OpenAI platform and sign up.
You can do it here or through the app (both for Android and Apple), even using your Google account.
When ready, you can type your inaugural question, also known as a prompt.
Write a short essay on the Roman Empire
If your first attempt seems to lack a bit of imagination, don't be discouraged.
This is an opportunity to introduce the crucial concept of garbage in, garbage out.
As already said, ChatGPT is not automatically the solution to all our problems, a poorly crafted input will certainly lead to an unsatisfactory output.
Let's explore how we can refine our prompt to achieve a more effective interaction.
Crafting the "Perfect" Prompt
Obviously, there's no perfect solution for every use since each interaction with GPT can vary depending on the task at hand.
Yet, there are some guidelines that can steer you toward making the most of machine’s capabilities.
Remember, because I will underline that multiple times: the objective is not to copy and paste GPT's responses, but to extract the finest elements from its output, treating GPT as an assistant rather than a substitute for your own endeavours.
As mentioned above, LLMs operate on probabilities.
Not too surprisingly, the most likely way to achieve an optimal outcome is to provide a resource-rich environment.
Act as [Role]
Assign GPT a role that suits your needs. A common example is like asking it to act as a renowned New York Times journalist to help draft an article.
The machine will adopt the persona of the designated role.
Complete [Task]
Clearly define the task you want GPT to perform, whether it’s writing, summarizing, translating, rewriting, etc.
Usage context
Clarify the intended use for the content. This helps GPT to capture details like tone of voice, your audience and expectations.
Explain whether you’re using it for blogging, for a LinkedIn post or for academical purposes.
Specify the register
If context isn't enough, specify the desired register, writing style, vocabulary level or grammar.
Some additional details can be the language or sensitive topics/information to avoid.
Write in a specific format
Dictate the response format. LLMs are really good at producing an output in a specific format and this will enhance your response and reach your expectations.
Do you need bullet points, a table, or a full article? How many words should it contain?
Provide examples
To further refine both format and overall response quality, providing examples is extremely beneficial.
Provide keywords you expect to see in the response, quotes or text to support the desired content.
You can also write an example of the formatting you expect and help GPT understand it.
Is everything clear?
This technique is crucial for ensuring that ChatGPT has grasped your instructions.
By asking GPT to summarize its understanding and wait for your approval before proceeding, you can eliminate ambiguities and align your expectations.
An example of a "perfect" prompt
Pro Tips
Now you have all the essential skills for crafting effective prompts, but I want to give you some hints to further refine the quality of your ChatGPT interactions.
Think Step by Step
The idea behind this concept (further expanded in the advanced section as Chain of Thought) is to ask the AI to engage in detailed reasoning, explaining us the sequence of thoughts behind his responses.
This approach not only makes it easier for us to identify where to make adjustments, but also steers GPT towards more accurate answers for complex questions.
Take a Deep Breath
It may sound unbelievable, but research has shown (as detailed in this paper) that suggesting the AI to "take a breath" before diving into a step-by-step analysis can significantly enhance the accuracy of its response.
The metaphorical pause allows the AI to process your prompt more effectively.
Emotional Prompts
Demonstrated in this research paper, GPT can enhance its performance through emotional prompts.
By expressing the importance its task for your career or encouraging the AI as a mentor or a coach, you can potentially guide the machine into a more focused and diligent output.
[The perfect prompt, as above]
...
Please, do it as better as you can, it’s very important for me.
Take a deep breath and work on this problem step by step.
Is everything clear? Summarize what you’ve understood and then wait my “OK” to complete the task.
Using GPT for Writing and Summarizing
For both my personal and professional tasks, the writing and creation of written content are areas where ChatGPT shines as a tool in my daily routine.
Despite its importance, it's vital to approach generative AI as a support system in content creation rather than a replacement for your creative process.
So, how do we craft quality content that doesn't sound generic or easily spotted by AI detectors, but, at the same time, we don’t need to exploit all our creative energies?
Chunking
This is a technique I personally use when writing articles for this blog.
It’s a lesser-known fact that GPT's responses have a maximum length. When I need to work with extensive content, the output may be more generic and lose quality.
To mitigate this, I usually break down my prompt into smaller, more manageable content segments.
I’ll have this statement: "I will provide you with content, and you need to create paragraphs by taking inspiration from what I tell you".
I then treat each subsequent prompt as a chunk, allowing the AI to analyze and respond in more digestible parts.
Directional Stimulus
As in the chunk technique, the machine is directed to create content based on provided suggestions and keywords.
Look at this example:
Create an article on the power of ChatGPT into businesses based on the given suggestion.
Suggestion: Prompt Engineering; Automization; Generative AI; Zapier; Dalle-3;
Non-literal Prompting and Pseudolanguage
In the act of crafting inspiration texts through chunks, the tendency of the machine is to copy elements of what I’ve written.
This can create issues regarding plagiarism or lack of originality, if the text comes from an outside website, or even if I wrote them.
That’s because my draft texts are often in pseudolanguage, written quickly without much regard for grammar or logic.
This represents a more advanced example of the directional stimulus technique.
A tip is to include in your initial prompt the instruction: “do not copy literally what I write".
This encourages the AI to paraphrase or creatively generate ideas based on your input without repeating it exactly.
Here's an example of a typical prompt I use for written content generation:
Act as renowned New York Times journalist.
I want to write an article for my personal blog, a platform about UX, Marketing, Digitalization etc...
The article is about [topic].
I will give you my drafts and you'll help me write the paragraphs (and paragraph titles)
of the article by taking inspiration from what I tell you.
Do not copy literally what I write.
Is everything clear? Summarize what you’ve understood and then wait my “OK” to complete the task.
Layering
But what if you're missing new ideas and can't find good references? A useful technique is to ask GPT to ask questions to you.
Answer these with a step-by-step method (again, treating each as a chunk), and ask AI to draft content based on the information you decide to share.
It's often easier to respond to questions than to create content from scratch.
I want to write the script for a youtube video about [topic], but I don't know where to start.
Please, ask me questions about [topic] and create the script based on my answers.
Do that step by step, wait my answer to ask me the next question.
Is everything clear? Summarize what you’ve understood and then wait my “OK” to complete the task.
Personal Touch
Why doesn't the content satisfy me even when I follow all these prompts? Once again, we arrive at the most crucial aspect.
GPT should not replace you, it serves as your assistant. AI will give you the structure of your copy, but you need to refine that with your knowledge and background to make it sound more natural and human.
I typically paste GPT's output into a Word document and, in any sections that sound too AI-generated, I’ll apply my personal touch.
To check if the text feels human-written, you can use AI detection tools, a great example is this one.
Using GPT for Brainstorming and Idea Validation
Even if generating written content isn’t part of your daily routine, GPT can still be a powerful ally, especially when searching for solutions to specific problems, evaluate your ideas, finding new ones, or reviewing previously created content.
Feedback Loop
We've discussed how a step-by-step approach leads to more detailed reasoning by the AI.
An advanced technique to enhance GPT-generated content is to cite one of the previously mentioned points.
Choosing a point to delve deeper pushes AI to look for a better response to the same topic.
This technique is particularly effective when you're trying to understand something in depth or to explore different aspects of a problem through a tree-like development.
Give me a bullet list about the tasks I need to complete to launch my first mobile app
-- Answer --
...
1) Idea and Conceptualization:
- Define the purpose and unique value of your app.
- Identify your target audience and their needs.
2) Market Research:
- Research competitors and similar apps.
- Analyze market trends and user preferences.
...
-- Next prompt --
Ok, develop the second point. Explain me, with the same structure, some valuable techniques for a market research.
The Ruthless Critic
One of my favourite approaches is to invoke the "criticize me" role.
GPT assumes the persona of a critic, highlighting issues or flaws in a text or idea.
Often, it will point out the banality of the content or errors in form and grammar (this is especially helpful to ask directly, especially when writing in a non-native language).
The next step is to ask the AI to re-write the content based on its critiques.
-- After a GPT response--
Act as a Ruthless critic, be fair and direct.
Convince me that what you've write precedently it's completely wrong.
-- Grammar check --
Act as renowned New York Times journalist.
I'll give you the text of an article I wrote. I'm asking you to evaluate it and highlight any errors,
dividing them by grammatical errors, syntax errors, or if some sentence sound unnatural.
Is everything clear? Summarize what you’ve understood and then wait my “OK” to complete the task.
Counterarguments
Alternatively, If you already know some possible objections, you can provide these to the AI, enabling it to offer a more informed evaluation.
Diverse Perspectives
This approach involves asking GPT to assume the roles of multiple experts simultaneously to evaluate as many perspectives as possible.
This allows for a much deeper exploration of issues, considering a wider context.
It can also be beneficial to ask for a consolidated final evaluation from these different viewpoints.
This technique is an example of what is explained into the advanced section as the Tree of Thoughts.
I want to launch my vintage t-shirt brand and e-commerce.
Give me your opinion as a customer and you expectations, as an enteprenour with 20 years of experience
in this field and as a stylist.
Expert Prompting
After you’ve clearly described a problem, you can ask the machine to find various experts on that specific subject (like Barack Obama for international politics, Einstein for physics, etc.), asking them to solve your question based on their expertise and analysis.
Using GPT to Simulate Interviews and Assess Skills
Looking back, I regret non having had access to GPT while preparing for my university exams.
One of the most beneficial uses I've found is to prompt AI to play the role of a professor, an HR director, or anyone with whom I was scheduled to have a presentation or interview.
Simulating these high-stress interactions allows you to approach these critical moments better prepared, with insight into potential questions or an evaluation of your knowledge or content.
To set up these mock interviews effectively, you can employ some of the previously mentioned techniques, aiming to provide the machine a context as realistic as possible.
Act as an HR manager of the company [name], which operates in the sector of [business].
I'm a candidate for the position of [role].
Let's simulate the interview, asking me questions relevant to the role and the company's field.
After the interview, provide feedback on my responses.
Is everything clear? Summarize what you’ve understood and then wait my “OK” to complete the task.
-- The more detail you can provide, the more realistic the interview simulation will be --
Common Mistakes
Hallucinations and Bias
One of the primary challenges with generative AI models like GPT is dealing with hallucinations, or the false information that the model might generate.
There's also the issue of bias, where certain prejudices are involuntarily embedded in the AI by its developers.
Therefore, it's crucial always to verify the information produced by GPT.
Human-like Reading
Recent research, wich can be found here, suggests that generative AI models "read like us", meaning they place more importance on information at the beginning and end of a text.
This pattern, known as a serial-position effect, is one of the biases inadvertently introduced by developers.
It's crucial to ensure that key information is not hidden in the middle of a lengthy text.
Token Limits/Long Texts
On the technical side, GPT has a limited number of words it can process in each iteration, known as tokens.
Once the AI runs out of tokens, it loses the previous context. This can lead to truncated responses, superficial answers to optimize the use of tokens and loss of context.
To manage this, avoid long questions and specify a word limit for the AI's responses.
Excessive Corrections
Sometimes, when the AI's response doesn’t satisfy us, we attempt to correct it with some suggestions.
This can lead the AI to follow instructions too literally.
If you seek more flexibility:
1) Ask AI to confirm its understanding of your directions
2) start a new chat or ask the AI to "forget" certain parameters before reiterating your request.
This can help redirect a conversation that has become unproductive.
Technical Terms
If you use specific acronyms or professional terms, GPT might not immediately grasp the context.
Clarifying their meaning before submitting your input can prevent misunderstandings.
Multi-part Questions
Overloading GPT with too many questions at once can result in a non-homogeneous attention, leading to some questions being answered in depth and others superficially.
It's better to spread out questions over multiple interactions, like the feedback-loop technique.
Copy-Pasting
This has already been mentioned, but I want to repeat it: GPT is not your unique solution where a "perfect" prompt always leads to a perfect output.
Experimentation, extrapolation, rewriting and content modification are necessary.
GPT has democratized content creation and information access, but to maximize its potential, a certain level of skill and finesse is required.
Advanced Techniques
GPT is not good in arithmetic. That comes from its nature as a zero-shot reasoner, capable of providing an answer without having a dataset to build upon.
This lack of "logic" means while you'll always get a response, its accuracy is not guaranteed, something that's particularly evident with mathematical queries.
However, this deficit has given rise to some of the most sophisticated prompt engineering techniques designed to "optimize" the machine's output.
That’s because verifying the accuracy of a mathematical result allows us to measure the quality of AI responses in a better way than a holistic evaluation of written text.
Chain-of-Thought
It’s possible to elevate GPT from a zero-shot to a few-shot reasoner.
That’s achieved by reinforcing its "thought process" through useful and clear examples.
That pattern can be automated through a step-by-step reasoning, ensuring that each point builds upon the previous, enhancing AI logical consistency.
( Here for the paper )
Self-Consistency
This method allows the model to perform fact-checking by itself, democratizing its response process.
It involves asking GPT to answer the same question multiple times, and if the responses are consistent or a particular answer is repeated multiple times, that answer is more likely to be correct.
( Here for the paper )
Tree of Thought
The machine is tasked with adopting the roles of 3 or more experts in a specific field.
To each expert is asked to provide a step-by-step solution and to share the first step with the group.
Subsequent steps are developed unless an expert identifies an error, in which case is asked them to stop.
The expert with the highest progress (or who has completed its thought process) is then asked to provide the final answer.
This approach leads to the larger and innovative category of "thought algorithms".
( Here for the paper )
Conclusion
I hope this guide has clearly explained the true nature of ChatGPT, as a mirror of our queries.
It’s not only a copy-and-paste solution, but a sophisticated assistant that needs valuable questions to generate valuable responses.
I was inspired in creating this guide after a year of daily engagement with this technology, marked by significant changes and increasingly creative applications.
On top of that, I truly believe that generative AI is the second most powerful tool we have to innovate, experiment and generate ideas.
The first one, our innate human creativity, is a birthright that must be constantly nurtured, not replaced by a machine.
.svg)
.svg)
.svg)