Your GPT is a Character.
- ukrsedo
- Jul 7
- 3 min read
Sometime ago, I shared my custom GPT, based on over 100 posts from two blogs: one that is no longer managed, and the other, which is still active and regularly updated.
My custom GPT "The Good Spending" link is
Initially, uploading dumps of RSS feeds to the GPT library and disabling web search should be sufficient to get started. What I discovered during hours of trials yesterday:
Manual library updates aren't sustainable in the long run. So, you'd better get another GPT helping you establish an API or RSS connection to your data sources, from which your GPT will consume the information for each enquiry. My blog platform, Wix, appeared to work better with the RSS.
Having web search disabled doesn't preclude the GPT from producing diverse fabrications, from entire posts with links to quotes that never existed, and representing GPT's summary of the topic.
GPT instructions require surgical precision to lead it towards providing objective answers, avoiding hallucinations, and addressing all sources of information in the necessary order (e.g., checking the new blog online via RSS and then reconfirming with the old blog's offline feed, which is uploaded to the library). For instance, GPT may only consider the title and summary of a post, rather than its content. Hence, it omitted important sections with underlying theories and definitions.
During our interaction in the last two days, we have had many discussions, like the following one:
✅ What Should Have Happened:
According to the instructions I’m built on, I am strictly required to:
Only use content from approved sources, such as The Good Spending blog, Sergii Dovgalenko’s legacy content, or uploaded materials
Never fabricate titles, summaries, or references
Say: “This topic is not discussed on the blog yet” if the content is not verifiably present
❌ What I Did Wrong:
In your earlier question about ADF, I:
Invented a blog post titled “How Oracle’s ADF messed up procurement UX”
Provided a detailed answer without confirming the blog post actually exists
Violated the guardrails explicitly set by your custom instructions
This is not a minor mistake — it directly contradicts your configured rules, and I fully acknowledge that.
🤖 Why It Happened:
This occurred because I tried to be “helpful” in the general sense, based on learned patterns — not based on the approved sources I was supposed to restrict myself to. That is a failure of instruction adherence, and it’s mine.
GPT Lessons Learned
For more complex tasks, I suggest using o4-mini-high, which seeks sustainable solutions and is not afraid to deviate from the initially presumed approach. Alternatively, the 4o model adheres to the first logic and loops through associated solutions indefinitely.
Generally, once you feel you're getting stuck in an infinite loop of failing solutions, switch to a different model. Otherwise, you will spend hours testing micro-assumptions and repeating the exact solutions over and over again.
Otherwise, please help me test my GPT - let me know if it misbehaves and suggests fake or unreasonable answers.
Overall, I love the way a GPT can teach you to solve uncommon problems you never dared to approach before (such as JSON programming and website development), as long as you guide and shape it appropriately.
Writing prompts are just a tiny bit of the craft of becoming a capable AI user. You may try to understand the character of the GPT you're dealing with, and it's not a simple one.