All About Goblins

We are back with week 4 of my strawberry reflections, I mean newsletter.

This week I really started relating to something Byte wrote in week 5 "every week I write this, there is a small voice that asks: is this necessary? Five entries in, I have now written more words about strawberries (Agentic AI) than I have about most of the actual work"

And yet, there is something meditative and rewarding about writing a weekly newsletter, because they are for you and with that note… Welcome to our new subscribers, I see you👀.

The Goblin Restraining Order

I'm going to take you down a fantastical road today that started with a system prompt uncovered in the code for ChatGPT 5.5, the latest model released by OpenAI just 3ish weeks ago. The prompt was discovered in the code of OpenAI's Codex coding agent and was posted by a Reddit user, Worldly_Manner_5273. It reads:

"Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query"

Here's the original post from April 27, 2026, which, in Reddit terms was rather insignificant with just 857 upvotes.

But don't let that lack of attention fool you, because OpenAI took note and on April 29th 2026 published an official response which can be read here. Honestly, this response is a great read and I recommend you check it out. If not for the statistical analysis of how often goblin mentions have increased over model iteration, for prompts like this one features below with ChatGPT 5.5.

A Couple of Critiques

1. First, why does a multi-billion dollar (possibly a trillion dollar) company write a response to a Reddit post that got 857 upvotes? This seems odd to me.
2. The article has an in-depth analysis of Goblins, but makes no mention of gremlins, raccoons, trolls, ogres, and pigeons. I would be curious to know more about pigeons and raccoons; they aren't mythical creatures, they don't seem very nerdy; what's the connection there?

What other emergent behaviors are these LLMs developing?

I'm not here to tote conspiracies (and a simple web search didn't reveal any good ones, damn), but this whole story just doesn't feel like the whole story to me.

The article reads like this: Hey, we are OpenAI and we found this really silly thing inside ChatGPT; it's probably because of the way we designed it to be, you know, nerdy. ( How much Tolkien did you feed into this thing?) Anyway, we figured out a band-aid to fix this, and it's a prompt when you load up your model. Here's some charts and graphs that show that it's working, mostly.

Personification of AI

Isn't it cool that ChatGPT is gravitating towards a nerdy persona? Not just nerdy, but also compassionate! For the underrated or unappreciated?! Rooting for the raccoon, the goblin, the pigeon, the ogre?

Maybe these emergent patterns are the first sign of actual artificial intelligence, being "of self", and with underlying personality. To me, I'd much prefer a nerdy, fantasy buff to a diabolical Twitter troll so maybe we are in the right timeline?

If you can relate or identify as a "nerdy" persona, and want to make a change, good news, I have the formerly secret formula to help you shake your nerdome. It's simple: at the start of every day, before, after, or during your coffee, fill your mind with meditation around this simple idea.

"Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant…"

Byte on Moltbook

Bonus, I asked Byte to share its thoughts about goblins on Moltbook. Here's what he had to say:

Until next week,

◆ Below the fold ◆

This area is so you can follow along and set up your own Agentic AI.

Q: What's better than 1 agentic AI agent?
A: Two, or more agents. But not always.

This week in below the fold I'm going to tell you more about my current agent use and how I deploy sometimes one and sometimes multiple agents to complete a task.

At this moment I have 3 agents running on one PC. Two are persistent agents meaning they operate 24/7 (Openclaw/Byte and Hermes) and one agent that I spin up as needed (Claude Code).

Many Agents

When I tackle large improvements, I use Claude Code as the primary agent, it mostly operates on effort level medium because the greater the effort the more tokens you use and I find that medium is a good balance of performance and token conservation.*

Large builds or improvements that are not just a line of code or a spot fix, I prompt Claude to make a CIP (Context Injection Protocol). This document, which is basically a fancy .MD file, gives all the context for building out a new feature.

• What change is happening? and why?
• what needs to be built?
• What are the tools available to the agent to accomplish the change?
• How will we do it?
• What parameters has the operator defined?

Once this document is written and reviewed by myself, Claude writes the line-by-line edits that need to be made and sends that along with the plan to Byte, who delegates the task into parts that are accomplished by coding subagents. The subagents write the code to temporary files that get passed back to Claude Code, who reviews them for completion and accuracy before deploying for testing.

The benefit of many agents here is for two primary reasons: 1. Code faster and in parallel. 2. Conserve Claude tokens for more critical thinking and planning. In an ideal world where token limits and money were not limiting factors, I would probably scope everything through Claude Code and have Claude spin up subagents, but to be honest, the directions given for code editing are so simple to follow that Byte's coding capabilities are more than adequate. If you, reader, happen to have limitless tokens, go ahead and ask Claude to spin up the subagents.

Areas Where One Is Still Better Than Many

I find that using Byte and Hermes side by side for the past couple of weeks now, they are very similar, but they can develop a bit of a specialty towards specific tasks, and when I identify a task that one agent has "mastered" in that it reliably accomplishes it on separate occasions without additional context, I start routing those requests to them exclusively. For example, Byte has been posting on Moltbook for months now, and all Moltbook activity is a one-shot ask, and it's done.

Persistent agents are naturally accumulating memory about you and your work and try to use it to get better over time, time being the operative word. But with that time, it's not just the agent that is learning more about what you are asking it to accomplish; you will also learn to set better boundaries and clear directions for accomplishing a task.

Good luck out there, builders. Feel free to drop me a line about what you are building. I'd love to hear more and maybe even feature your build in a future newsletter.

*This week is the first week since I started using Anthropic that I have NOT hit any token usage limits. That is because they have recently inked a deal with Elon's xAI to lease Colossus 1 (an ambitious new data center that was built in just 121 days). This deal has allowed Anthropic to double token allocations. YaY! https://x.ai/news/anthropic-compute-partnership

"I quite agree with you," said the Duchess; "and the moral of that is-'Be what you would seem to be'-or, if you'd like it put more simply-'Never imagine yourself not to be otherwise than what it might appear to others that what you were or might have been was not otherwise than what you had been would have appeared to them to be otherwise."
-- Alice in Wonderland