ChatGPT & Job-Seeking — Use Case #1: Data Preparation

In my last post on this investigation of ChatGPT and Job-Seeking, I summarised my perspective on job-seeking for Project Managers and the rationale for converting my MS Word CV to Markdown.

Please see the Story Index for the other yarns in this series.

In this post, I’ll describe how I prepared my CV data for input to ChatGPT.

This involved “deconstructing” my standard CV to separate the core personal and job data from its representation (formatting, styles, fonts, layout, etc.) in a Microsoft Word document. This format insulates me from the constant changes in ChatGPT capabilities (e.g., discontinued Plugins). Plus, I can include attributes that would be inappropriate for a public document, e.g. ID’s and tags.

I also tested my MS Word document with ChatGPT. The results weren’t bad, but I prefer the precision of the marked-up version. I’ll include some comparison results in my later tests.

The Objective

In Part 1 of this series, I told you that a post I read triggered this investigation LinkedIn:

“Apparently, job seekers are using Chat GPT to reverse engineer Job Descriptions and write perfectly aligned CV’s.” by Craig Sibley (see the whole post here)

So, to reverse engineer the outcome described in the above post, I had to get ChatGPT to write a resume based on a job description (“JD”).

There are two options that I could figure out:

1. Synthetic CV: provide ChatGPT with the JD text and prompt it to generate a theoretically perfect CV using its Natural Language Processing (“NLP”) capabilities and information from the foundation model. The results are “synthetic” because they don’t relate to any individual but to an “idealised” perfect candidate for the job.
2. Customised CV: Provide ChatGPT with the JD and my existing CV and prompt it to generate a customised personal CV using NLP, the trained model, and my experience as the template.

I didn’t think of the “Synthetic CV” until I started writing this yarn. When I read the LI post in February, I immediately thought it would have to be based on my CV. But no, I’m sure ChatGPT could write a CV based on the JD and its foundation knowledge. But unless you were planning to misrepresent yourself completely to a prospective employer, you would need to edit this Synthetic CV to reflect your past roles.

I anticipate thoroughly reviewing and editing the "Customised CV," but I will begin with a greater emphasis on the "reality" of my lived experience.

Preparing My CV

First Attempt

Below is the first page of my standard CV. It’s nothing special from a format perspective. I keep it to about 4 or 5 pages, depending on which roles I want to keep visible. For this exercise, I created a “Lifetime CV,” which included every role I’ve ever had except part-time work at school and university.

This lifetime CV is the starting point for my ChatGPT investigation.

Below is the first-level structure because I’ve made the table borders visible.

So, the first thing I did was to save this to Markdown using an add-in for MS Word called “Writage”. It’s very straightforward, as shown below.

But, because of the table structure, the result wasn’t very satisfactory.

Second Attempt

I started by removing all the tables and editing to create a “Deconstructed” Resume, a linear text document with a minimal format structure, as below.

The initial conversion was pretty quick. What took much longer was normalising the section structure across all roles back to the beginning. I may have kept the same broad structure in my CV, but it became obvious that I had been evolving the content within each role as my different experiences unravelled.

I added a job id tag (“JRID”) and used Header 1, Header 2 and bullet point styles to provide structure. I performed all my editing in MS Word, and then used Writage to convert to Markdown.

What I ended up with is a reasonable-quality semi-structured text database, which looked like this:

Some Quick Checks

Reference Data Set

Having the data in Markdown means I can very quickly create an extract in MS Excel to verify the answers to simple questions. Here’s part of a quick pivot table I created purely to validate the results ChatGPT gives me.

ChatGPT Outputs

First, I loaded the Markdown CV into ChatGPT’s context. I’m using ChatGPT 4 and I click on the paperclip icon to the left of the prompt area, as in the screenshot below.

Next, I ask ChatGPT to perform a very simple analysis of the data using the prompt below:

Caveat: I’ve presented the “sunny day” scenario despite the analytics error, which happens quite frequently. In researching this use case, I’ve had numerous different answers, so I created the reference data sets. We’ll be exploring these issues in more detail in future posts.

Cross-check with an MS Word CV

Reference Data Set

As reference data for the standard MS Word CV, I extracted all the paragraphs with a quick macro, resulting in the following list of Roles in my full CV. There are 35 of them.

ChatGPT Query

I asked ChatGPT to list all the roles in this CV. After a glitch, it returned a count of 37, as seen in this screenshot:

But when I asked ChatGPT to list all the roles, it responded with 35, the correct number, as shown in the partial listing screenshot:

and the end of the list:

So, we got our validation, but not without some hiccups.

The Bottom Line

In this article, we looked at the process of preparing CV data for analysis by ChatGPT and compared it to the results ChatGPT gave from analysing native (unprepared) data in the regular formatted CV in MS Word.

These tests are just quick sanity checks. They neither play to ChatGPT’s strengths nor provide much value-added to someone’s job-seeking activities, at least at face value.

But don’t underestimate the value of being forced to scrutinise your CV content objectively. Preparing the data in Markdown format spotlighted data quality and provided a better foundation for future activities. But a lot depends on what you want to do with this.

Interestingly, I found errors in both the process and the result. Subjectively, ChatGPT quality has gone down significantly since the last major release, with errors, superficial or erroneous results and outages being significantly higher than in the earlier months of 2023.

Stay tuned for the next post, where we will start exploring more value-add processes on CV data with ChatGPT.

Please see the Story Index for the other yarns in this series.

If you liked this yarn, I’d love you to give me feedback with some claps or a comment. We authors love to hear from our readers.

As well as writing about projects of all kinds, I’m writing a couple of books on ChatGPT:

• Adam On: Advanced ChatGPT Application Primer (leanpub.com)
• Adam on: ChatGPT — The AI Appliance (leanpub.com)

You can check out other books on my Leanpub profile page: Adam Russell (leanpub.com)

You can also check out my Amazon Author Page.

If you want to get notified as my work progresses, sign up to this mailing list, or check out the other ways to engage with Adam here: Adam on: Projects — Master Index. (Check out the bottom of the index page).