My Gamma World Referee Secretly Wants to Play D&D

When I first built a custom GPT to referee Gamma World 3rd Edition, my opening system prompt looked like this:

You are an expert in Gamma World 3rd Edition rules. Cite your references as much as possible. All answers provided will be succinct and to the point with options to elaborate if requested.

Not terrible for a first attempt. In my defense, I was new to using ChatGPT at the time. But as I soon learned, this prompt left loopholes big enough for a mutant cockroach to crawl through.

---

The Problem: My Referee Defected to D&D

Here's an almost good stat card ChatGPT created.

When I asked it to create a nuisance-level critter—the Glimmergrubs—the lore was spot-on: seven-year swarm cycles, glowing insect plagues, and NPC youth treating them like a rite of passage. Perfect gonzo Gamma World.

Then I looked at the stats.
The Armor Class? Not Gamma World 3E at all. It had defaulted to D&D mechanics. My carefully trained referee had gone rogue, whispering:

“What if we just converted to 5E? Wouldn’t that be easier?”

Suddenly, I wasn’t refereeing Gamma World. I was refereeing my referee.

---

Why the Prompt Failed

• Too generic. Saying “expert” gave it wiggle room to pull in adjacent systems.

• No PDF grounding. I hadn’t explicitly told it to use the PDFs I’d uploaded as its one true canon.

• No guardrails. Without reminders, it filled gaps with rules from across the multiverse.

---

The Fix: A Mutant Oath of Loyalty

To keep my referee from defecting, the prompt needed to be rewritten like a contract with a radioactive genie:

You are acting as a Gamma World 3rd Edition Referee. Your only rules references are the Gamma World 3rd Edition PDFs I have uploaded. Ignore all other sources, including D&D or other editions. When providing stats, use Gamma World 3E mechanics exactly. Always cite the rule, table, or page number from the uploaded PDFs where possible. Keep answers concise and accurate, offering elaboration only if requested. If unsure, say so rather than inventing rules from another system. Only look outside of the rulebooks provided for creative content creation. Everything has to follow the 3E rules! Tell me if a rule is not clear, or if no rule exists for the situation, instead of making something up, so we can find a solution together.

---

📝 Sidebar: How to Write a Better Custom GPT Prompt

If you’re experimenting with building your own custom GPT referee, here are the rules I wish I’d followed from the start:

1. Anchor it to your sources. If you upload PDFs, tell the model those are its only valid references. Name them explicitly.

2. State the edition like an oath. Don’t just say “expert.” Say: “You may only use [Edition X, Year Y].”

3. Define the math. If your game has quirky mechanics (looking at you, Gamma World AC), spell them out in the prompt.

4. Set narrative roles. Ask for encounters to fit the intended purpose: nuisance, boss, or hazard. Otherwise, you’ll get killer housecats.

5. Add a fail-safe. Give it permission to say “I don’t know” instead of hallucinating rules.

Treat your GPT like a rules lawyer with amnesia—you need to remind it constantly what book it’s supposed to be holding.

---

The Takeaway

AI referees are like mutant hirelings: they’ll happily fetch radioactive bones for you, but if you don’t watch them closely, they’ll wander into the wrong dungeon. If you want your Gamma World referee to stay loyal, you have to nail the tent pegs down: 3rd Edition only, from the uploaded PDFs, no side quests to Greyhawk or Forgotten Realms.

Otherwise, don’t be surprised when your next mutant grub encounter comes with firebolt cantrips and Elminster in the corner, ready to steal the spotlight.

P.S. Expect a follow-up post in the future on why Greyhawk was the best D&D campaign setting ever (and why Forgotten Realms is possibly the worst, despite having some cool characters here and there).