I swear I am giving ChatGPT Dementia

And no matter how you use AI, you must always do the following:

• Be very specific in what you ask/tell it to do.
  If you provide general directions, most often in technical situations you will not get what you wanted. AI, with its vast technical knowledge, will chase what I call the “bright and shiny.” If it doesn’t chase the bright and shiny (which is usually overkill and requires more work than is needed), it just won’t be on the same thought process you are on.
• Always, and I mean always, review what the AI provides you.
• Be very aware that the AI you are using has resource limitations. So if you are in long discussions, these resource limitations can cause the AI to do what I call drift.

So, why am I ranting, you ask? A number of reasons, to be honest.

I have learned the hard way about the resource limitations. As I encounter them, I try to find ways to address them so I can continue the work I’m doing.

I swear to God, every time they update the model that AI uses… I may not be back to square one learning how to work with the limitations that my lower-paid subscription has… but there always seems to be a new twist: more limitations, different behavior, or new restrictions. One step forward, two steps back.

No shit. AI has good days and it has bad days. There are days I can pound out hardcore code with lots of interaction, and there are other days AI just sucks.

The responses are garbled. I have to do hard refreshes on the browser window to actually see what the hell it is saying. The hard refresh clears up the garbled response.

It doesn’t matter some days that from the very beginning, even before we start hardcore coding, the AI just is not being helpful.

Then you have the days where you’re being productive with AI, and then all at once everything goes south. It’s like… WTF, Gadget.

I’ve learned that I have had to develop guard rails, or session rules, that AI should follow for the current session. I make AI post those rules at the very beginning, and it tells me it will follow them once it posts them. That statement is bullshit. Yes, it has helped a lot of times. But damn… it just gets ignored other times.

When I say guard rails or session rules, how much more explicit do I have to be than this?

Mode: Strict pair-programmer
No advice. No extras.
Only output what I asked for.
If info is missing, ask (don’t guess).

✅ Workflow rules
One file at a time
Keep changes small and reversible
No scope creep
Don’t “clean up” unrelated code

✅ Architecture / stability rules
Create helpers/services when logic repeats
Maintain a single source of truth
Prioritize portability (don’t hard-wire environment stuff)

✅ Build discipline
After each change: build and fix errors
Verify via output/debug
Stop only after a clean compile

I mean, seriously. GADGET, these are the rules. Follow them.

Now, the reason I have to rant is that I have MONTHS of work into my technical project. This project is a personal Windows password vault that is built around the following framework. MY framework.

• Layered security
• Multiple levels of encryption
• Ease of use by the user. In other words, the user interface has to be simple for anyone.
• Expected results should just happen.

The security and encryption layers are almost anal in how I have designed it. Yes, it can be hacked. Everything can be hacked. But sucker, you are going to bust your balls hacking it.

I am now seeing the light at the end of the tunnel of MY creation. And now, for whatever reason, AI has become more unstable than it was previously.

When I say gaurd rails or session rules. How more explicit do I have to be then this?

Mode: Strict pair-programmer

• No advice. No extras.
• Only output what you asked for.
• If info is missing, ask (don’t guess).

✅ Workflow rules

• One file at a time
• Keep changes small + reversible
• No scope creep
• Don’t “clean up” unrelated code

✅ Architecture / stability rules

• Create helpers/services when logic repeats
• Maintain a single source of truth
• Prioritize portability (don’t hard-wire environment stuff)

✅ Build discipline

• After each change: build + fix errors
• Verify via output/debug
• Stop only after a clean compile

I mean, seriously. GADGET, this is the rules. Follow them.

I know that AI is a tool. And honestly, my experience with it is mostly positive. I wouldn’t have the solid framework of this application that I have right now if I hadn’t used AI to get to the point I am now.

But right now, I am losing days of work because I just can’t always trust it when AI drifts on me in the middle of a coding session. Thank God I review everything, or this application would be in cluster-f**k land if I didn’t review everything.

Just to be honest, this drifting, as I call it, only occurs in my coding sessions. When it comes to other uses of AI (this blog, or other questions), it is fine. But this blog is not my core reason for using AI. The coding is.

Ok, I’m done ranting. Sorry for ranting. But we all do it. And I just have a blog to bitch and moan on when I feel the urge.