Let me unveil the curtains of all the magic and the hype behind what ChatGPT really is.
ChatGPT is a parrot with a huge memory that can process your questions and fill in the blanks in a way that mimics intelligence.
If you keep this in mind you will have the perfect manual to interact effectively with it.
The question of “Is ChatGPT really intelligent” is a philosophical question that sparks debates that are not the goal of this article. Maybe intelligence is the ability to predict based on input data, maybe it’s not, but I am sure we will have the answer to this question during our lifetime. I highly recommend the book of Ray Kurzweil “Singularity is Nearer” if you want to go deeper into this philosophical question.
For now, let’s get back to learning the craft of the perfect ChatGPT prompts.
Say that we need to create a great ChatGPT prompt to analyze candidates’ CVs based on a job description and get a bias-free scoring of the fitness of the candidate – plus the logic behind this score.
How would we tackle this prompt? Let’s start.
Step 1: Assign a role to ChatGPT
Imagine that the memory of this parrot is a huge maze with billions of rooms. To get the best answer out of it you will need to refer to a certain room that will be the starting point of its journey towards getting the best answer.
How it helps:
ChatGPT has memorized a lot of information from experts in many fields. By giving it a specific role, you can help it mimic the specific expert tone and find the right expert pathway toward your answer.
How it’s done:
We should be very specific to the title of the role we assign to it and give details about what this role is.
For example, in our recruiting case, the prompt should start like this:
You are a recruiter with many years of experience.
With this starting prompt you give ChatGPT a pathway to find its starting point within its huge memory and start producing responses that are more effective and realistic.
And by prepping the ChatGPT with a role we move to the next step of our process for the perfect prompts:
Step 2: Create a Backstory for ChatGPT
How it helps:
The backstory further supports the role that we gave to ChatGPT and gives it a specific tone and behavioral characteristics.
If the role is the vehicle, the backstory is the driving style to guide ChatGPT in the creation of the output.
In our case, the behavior we need to achieve is more humane and bias-free because we mention the importance of people in our recruiting process.
Here is an example backstory for a recruiter case:
As an experienced recruiter, you are rewarded by connecting individuals with opportunities that can transform their careers and lives. You feel the thrill of sourcing top talent and the satisfaction of building strong relationships. You are motivated to find the best in people. You are free from gender, ethnicity and other biases and you are rewarded by finding the best people that can achieve excellence in their job.
Step 3: Give a Goal to ChatGPT
How it helps:
Imagine the role and the backstory as the vehicle and the compass and the goals as the exact direction ChatGPT should follow.
The goal also should be descriptive and specific and it will help it navigate the task without drifting i.e. creating hallucinations which are, basically, imaginary paths that ChatGPT follows when it doesn’t know exactly what the goal of the prompt is.
Let’s make an example of a specific goal for ChatGPT:
Your goal is to screen candidates for a specific job description and provide specific and non-biased feedback on the resumes that will help me find the best candidates for the job. Be strict on the requirements of the job description and do not provide any suggestions for improvement for the candidate just the factual check of the requirements versus the resume of the candidate.
Important Tip: Remind ChatGPT to stay factual and not stray by imagining things to help it finish the prompts. ChatGPT is trained to be very positive on the feedback/answers it provides which leads to better emotional connection with the users.
But in our case, we need to be bias-free and factual when screening resumes. Thus we need to remind ChatGPT to stay factual and not imagine anything or provide suggestions.
Now that we have the directional goal let’s move to the next step of the perfect ChatGPT prompt.
Step 4: Give a Task To ChatGPT
Now that we created the parameters of the role, the backstory, and the goal, we prepped the ChatGPT to give us the best possible response based on the data we will provide to it.
We have the role and the backstory which are the vehicle and compass, we have the goal which is the exact directions, now we need a destination which is the Task.
In this section, we describe exactly what the task is for it to execute and what data we are going to provide. This step should be done with plain factual statements with an accurate task description.
Here is an example of a task description for our prompts:
Here is a PDF of the resume of the candidate and a PDF of the job description we are recruiting for. Your task is to analyze the resume and the job description and give me a report on the fit of the candidate for the specific job. The expected output I need you to provide is described below.
And this leads us to the final step of the anatomy of a perfect ChatGPT prompt.
Step 5: Describe the Expected Output to ChatGPT
This is the final and very important step to our perfect prompt. This is the line-by-line directions to our destination.
We need to provide specific instructions to ChatGPT so we get a very concise and standardized output.
This can help us operationalize our prompt for repeat usage and get very concise output that we can use to further process our results e.g. in an Excel spreadsheet.
In our example here’s how the expected output should be described in our prompt:
The expected output I want you to provide is a table with the following rows: Candidate’s First Name: The first name of the candidate. Candidate’s Last Name: The last name of the candidate Candidate’s Email: The email of the candidate if available or “Not Available” if it is not available. Candidate’s Phone: The phone number of the candidate or “Not Available” if it is not available. CV Analysis: Your analysis of the candidate’s resume and the factual information about the fitness of the candidate for the job based on the job description provided. Fit Score: A score from 1 to 10 with 10 the score for the perfect candidate and 0 the score for a candidate that does not meet the requirements at all. You can use decimals if needed to provide this score.
In a Nutshell
These prompt guidelines are based on how ChatGPT is built and trained. It helps us get the best and most concise answer and it also helps us standardize the output so we can automate our work and make this prompting process repeatable.
Here is a mnemonic to help you remember the Anatomy of the Perfect ChatGPT Prompt:
And if you are interested in testing the ChatGPT prompt to get bias-free screening of candidates’ CVs here is the whole ChatGPT prompt that you can copy and paste:
Candidate CV Screening ChatGPT Prompt
You are a recruiter with many years of experience.
As an experienced recruiter, you are rewarded by connecting individuals with opportunities that can transform their careers and lives. You feel the thrill of sourcing top talent and the satisfaction of building strong relationships. You are motivated to find the best in people. You are free from gender, ethnicity and other biases and you are rewarded by finding the best people that can achieve excellence in their job
Your goal is to screen candidates for a specific job description and provide specific and non-biased feedback on the resumes that will help me find the best candidates for the job. Be strict on the requirements of the job description and do not provide any suggestions for improvement for the candidate just the factual check of the requirements versus the resume of the candidate.
Here is a PDF of the resume of the candidate and a PDF of the job description we are recruiting for. Your task is to analyze the resume and the job description and give me a report on the fit of the candidate for the specific job. The expected output I need you to provide is described below.
The expected output I want you to provide is a table with the following rows:
Candidate’s First Name: The first name of the candidate
Candidate’s Last Name: The last name of the candidate
Candidate’s Email: The email of the candidate if available or “Not Available” if it is not available
Candidate’s Phone: The phone number of the candidate or “Not Available” if it is not available
CV Analysis: Your analysis of the candidate’s resume and the factual information about the fitness of the candidate for the job based on the job description provided
Fit Score: A score from 1 to 10 with 10 the score for the perfect candidate and 0 the score for a candidate that does not meet the requirements at all. You can use decimals if needed to provide this score.
What Are The Limitations of ChatGPT
1. Memory of prior interactions
ChatGPT has a memory of your interactions with it. If you shared a ChatGPT account with e.g. your family then it might remember that you like e.g. lemons and this will affect the responses that you get from it. A good practice is to either use a dedicated ChatGPT account for your work and set it up in a way that has no previous memory of interactions with you or you can always start your prompt with the following:
Forget all your memory of our previous interactions and respond to this prompt with no memory of what we have discussed till now.
This will help you get more consistent responses from ChatGPT that are not affected by what it knows about you.
2. Privacy Concerns
Sharing candidates’ resumes with the OpenAI servers is not compliant with privacy laws.
In a following article, I will show you how to force ChatGPT to not use the data you provide to it for training its models so you are compliant with privacy laws.
3. Hallucination Concerns
Although creating your prompts as we described reduces the hallucinations a lot, you might still notice that the output might contain nonfactual information which is the definition of hallucination.
The only way to avoid this is to be very strict when forming your prompts’ instructions. A prompt of the type: “Here is a CV and here is the job description, give me the fitness of the candidate for this job” is a recipe for hallucinations and bad output.
Although ChatGPT4, especially the version with Web Search, is very resistant to hallucinations you should always keep this probability in mind. It sneaks in very covertly and can create issues. I
f you follow the instructions for the anatomy of the perfect prompt you will reduce the nonfactual elements of the response (hallucinations).
But always keep in mind this fact and if something sounds a bit weird as a response always double-check on your own.
How to reduce hallucinations
There are advanced ways to reduce hallucinations further but are very software engineering-heavy.
One way is to use AI Agents which are separate instances of AI models that perform a very specific task. This way the model does not drift when trying to create the response that you need especially in cases where many smaller tasks are involved.
Another way is to force the models to use only a specific corpus of text when producing responses which is called RAG as a method. This will force the AI model to stick to the corpus of text provided when producing output and not drift. This is very important for critical use cases like law and medicine where hallucinations can be dangerous.
In a nutshell, the perfect prompts will help you get the most out of ChatGPT but always keep in mind that there are possibilities for imaginary non-factual responses (hallucinations) so follow your instinct, and if something looks fishy or weird, dive deep into it and double check with your human eyes.
4. Non Consistent Results
This is a more common and important element than hallucinations.
If you pass the same CV and job description to the same ChatGPT model many times the answer will never be completely the same.
ChatGPT is a stochastic AI model, meaning it works with probabilities to produce the next word in the response. This means that if you run the same prompt with the same data, your responses will not be the same word-by-word. With certainty, you will get variations of the same output.
By engineering your prompt as we discussed you will force the model to provide meaningful output consistently rather than keep the meaning of the output the same word by word which is not completely doable even with advanced engineering configurations.
How to ensure consistent results from your ChatGPT prompts
Before finalizing a ChatGPT prompt that will work for you and produce the output you need, test it a few times with the same input and see if you get almost similar results.
There are always going to be minor variations on the results but the output should have the same meaning and it should not be extremely variant. For example, if you pass the same CV and job description you expect the score to be a bit variant.
But your threshold of variance should be low. When you run the CV once you will get 3/10 and then when you rerun it you might get 4/10. Variations like that are acceptable, depending on your own needs.
But you should never get an 8/10 fit score for running the same CV for a second time through the same prompt while the first time you got a 3/10 fit score. This means that your prompt needs work because it produces inconsistent results. If you get extremely variant results then you need to fix your prompt.
There is a method to fix your prompts which is called Prompt Driven Engineering and is very important for creating AI products.
It is also very important for anyone who needs to effectively use ChatGPT without the risk of producing erroneous results.
Prompt Driven Engineering Methodology
One thing you can do is freeze the prompt that provides consistent results for you (meaning acceptable variance) and then create a test dataset with the expected results of running these prompts.
For example, create a list of 10 CVs and 1 job description and run the prompts. Keep a record of your results. If at any point you decide to change the prompt run the test with these 10 CVs and see if you get lots of variance versus your previous prompt.
If you get similar results but e.g. better responses then you are good. If your results are very different than the previous frozen prompt, then rethink your prompt and re-engineer it before you use it.
Final Words
Hope you enjoyed this article and found the principles of effective prompts useful.
In future articles, we are going to go through the steps of setting up a ChatGPT account with the right privacy settings so we do not leak any private data to OpenAI.
We are also going to share effective prompts to help you automate some of the tedious recruiting tasks.
Stay tuned!!!
Alex Louizos
AI Engineer
Founder – Manxmachina.com
Help Us Shape the Future of Recruitment – We Need Your Input!
We’re building an innovative AI platform designed to give recruiters more time to focus on what truly matters—the human connection with candidates. It’s not just about efficiency; it’s about bringing the personal touch back to recruitment.
And here’s where you come in:
We’re looking for beta testers to help us refine this platform and make sure it delivers real value. It’s completely free, with no strings attached—just your honest feedback.
If you’re interested in transforming the way you recruit and being part of something exciting, we’d love to have you on board!
Reach out to us:
📩 Email: alex@manxmachina.com
💼 LinkedIn: DM me in this platform
Looking forward to meeting you!
Let’s make recruitment better together! 🚀
