The initial setup for the CustomGPT showed some promise, I was really excited to test it out once I had the general setup done, but as predicted, my first round of testing revealed several areas that needs refinement, as it is just a concept build, to make sure it can actually work, now that I know it can run theoratically, I need to refine it.
Challenges with Archetype Generation
The archetypes generated by the GPT felt very vague and impersonal. The GPT struggled to capture the nuances of a user’s social behavior, making the results feel more like generic labels than personal insights.
I quickly realized that this was largely due to the absence of a structured framework to guide the GPT’s decision making process. Without a framework to follow, the GPT was left to making random decisions and asking random questions, which impacted the quality of the archetypes.
To address this, I plan to implement the below frameworks and see how it runs and tune it from there:
-Social Interaction Style (Based on the Extraversion Scale from Big Five):
-Social Role (Based of Leadership & Group Dynamics from Social Identity Theory)
-Conflict Handling Style (Based on Thomas-Kilmann Conflict Mode Instrument)
-Social Network & Engagement Context (Based on Social Circles & Network Theories)
-Emotional Intelligence (Based on Mayer-Salovey Emotion Intelligence Model)
Issues with Visualization:
Another issue that showed was with the visualization process of the archetypes. In some instances, the visualization process will not even trigger, and sometimes and generated images would contain game logos or elements that shouldn’t be in there, like random texts. Additionally, the visualizations themselves were somewhat generic, lacking strong contrasts between different personality types. For example, whether the archetype was a leader or a follower, the visuals don’t vary much.
Tuning Mental Rehearsal for Personalization:
The mental rehearsal component of the GPT is one of the most important aspects of this intervention. During my first test, the GPT guided me through a scenario that was highly personal, asking me to describe a real life situation in which I could apply my social strengths. That felt incredibly engaging for me as I could very vividly imagine myself in that scenario, walking around and interacting. However, in my later tests, the GPT defaulted to generic scenarios that felt much less personal, I feel like this diminished the impact of the exercise, as it no longer felt personal and is not tailored to my experiences.
I will need to tune the mental rehearsal section so that it is able to consistently replicate the personalized experience I had in the first test. This means ensuring that the GPT always asks the user to provide a scenario from their real life, rather than defaulting to a pre-scripted situation. This will involve asking the user about recent or upcoming social situations or challenges they have or might face and then guiding them through a mental walkthrough of that experience.
I want it to feel as personal and realistic as possible, to allow users to be able to genuinely mentally picture applying their online strengths in offline situations. If the users can fully picture themselves in these rehearsals, the chances or them successfully using these skills in real life will likely increase.
Next Steps
Moving forward, my primary focus will be on refining the archetypes and mental rehearsal sections, as these are the most important parts of the project. Both areas needs deeper personalization.
Oh and about the name of the CustomGPT, I originally named it the “Persona Explorer.” But let’s be honest, it’s a very generic name. I’ve taken to liberty to rename it “Jeff,” until it is polished enough to deserve another name. So for now, I will be working with Jeff.