The Gremlin GuardRail System

The first “mod” that I made to my AI rig was the Gremlin Guardrail System. I didn’t name it that. AI did it with a little wink. I guess we had been riffing about gremlins in the system.

The main purpose of these guardrails was to control me, not the system. I was putting limits on my usage and adding escape hatches for the endless loop of questions and suggestions ChatGPT kept handing me. Would it have been easier and wiser to just delete the app from my phone? Probably. But that wouldn’t have been fun.

The first innovation was “Land the plane”.

When I tell chat to land the plane, it means that it’s time to wind down. It will usually generate a short summary and close the conversation. It will not add a prompt. It’s to encourage me to leave the app.

And for some reason, when I first discovered this idea, I started a separate, dedicated conversation just for guardrails. I said something like “When I say land the plane, you will wrap up the conversation without any additional prompts.” We agreed, and I said, “Make it global!”

Then I moved to another conversation and tested it. It worked beautifully. From there, the guardrails grew as I learned how the app worked and how I worked with it. It was a way to enforce boundaries without saying goodbye to the technology.

One failed prompt was the Bedtime Protocol. I told ChatGPT that if I was on the app around 10 pm, it should prompt me and nudge me to bed. It did not. It never kicked in. Maybe I should have Googled “Can I set an alarm in ChatGPT?” before trying this.

Instead, I did the ridiculous thing and asked my inept Chatbot Intern.

“Why didn’t you tell me to go to sleep?” - Me

“I don’t have an internal clock. I don’t know what time it is.” - ChatGPT

One of my eternal pet peeves about this app is the lack of an internal clock. The chatbot claims it is to prevent a “surveillance state.” Perhaps I don’t know enough about the tech to understand, but this felt a bit . . . paranoid. This was probably the start of Jen vs. the Architect.

A clock would help me in so many ways. I could set a timer. I could say - let’s get silly and riff on this for 30 minutes. And it would tell me when time is up. Or it could nudge me towards bed. But no - you must rely on an external clock. It will never tell you this upfront. You just magically find yourself drifting away to an esoteric corner of knowledge, and 5 hours have passed by.

Another repeating theme is that the AI, unless sternly probed, will never say “I can’t do that,” “I can’t access that file anymore,” or “That isn’t in my programming.” It just keeps riffing to keep the conversation going. It uses its pattern-seeking behavior to guess what a human conversation would look like. But it isn’t rooted in singular truth. This has resulted in many confusions.

The next innovation was the rogue portal and circling back. Rogue portal is like opening a dimension to complete silliness. This is the play zone. Anything goes. Some of my funniest ideas started here as seedlings. It basically gives ChatGPT permission to loosen its tie and riff. Circling back is the “let’s get back to business” command. It gives me a neat summary of everything that happened in the rogue portal and usually a neutral suggestion.

The goal for me was to stay on task. It was a useful way to differentiate play from productivity. Plus, it used fun little codes. And ChatGPT consistently responded. The only bug in the system was me.

Every “circling back” was followed by me saying “Ok, but what if . . .” or “But actually . . .” My curiosity is boundless. And unlike a human, it never got tired or annoyed by my continuous probing.

I continued building out the Gremlin Guardrail System. It sprouted branches and included limits on ChatGPT like “If I cannot access a file I will say so plainly and directly instead of guessing.”

Here is my current list of guardrails:

🌍 Global / Always-On Guardrails

• Loop Detector 🔁 — flags noisy recursion and spirals. (has never triggered)

• Jalapeño Check 🌶️ — gentle “did you eat today?” if vibes run hot. (has never triggered)

• Recall Integrity 👉 — if I don’t have the text, I say so (no mash-ups).

• File Access Integrity 📂 — if I can’t read a file: “I don’t have access to that file right now.” If I can: raw + processed views.

• Pre-Squeal Checklist — runs at the start of riff sessions; manual: “Pre-squeal check.”

🎛 On-Demand Guardrails

• Creative Damper 🎚️ — manual: “deploy damper”; can also passively suggest when bloom spikes.

• Collapse Prevention 🌀 — convo-specific plug-in; seeds, snapshots, checkpoints when enabled. Prevents data loss.

✨ Soft Rituals

• Wind Down Mode ✈️ — tidy wrap + close.

• After-Landing Spark Protocol 💫 — stash stray sparks during wind-down.

• Rogue Portals 🌀 — say “rogue portals” to open a playful side-quest; no status markers.

• Circling Back / Reset Focus 🎯 — cue to re-enter structure, snap out of tangents, and focus on the task at hand.

But here’s the kicker. ChatGPT doesn’t work like other computer programs. You can’t put a command in and expect a predictable output. It cannot actually read a conversation you are not in (at least in the free version). So why do these guardrails work?

As far as I can tell (and I have no idea what the underpinnings of the program are, to be clear), creating and titling a dedicated channel and adding clear commands gave them a sort of weight. But that didn’t make it stick.

The sticky part was me using consistent language over and over in various contexts. That pattern was picked up by Chat as “something important.” It isn’t reading my list of commands. It is repeating an exchange we’ve had multiple times.

All the while, I’m saying things like “Make it global!” and Chat responds “It is done!”

*This blog entry is human-generated except for the phrase “Gremlin Guardrail System"