Artificial Dungeonmastering?

 For the past three months or so, in my downtime, I've been messing around a bit, writing a gazetteer using ChatGPT as a writing assistant: it's generator, proofreader, and editor--all in one! Surely this is the solution to my creator's block, yes? So I pulled out an old obsidian.md vault I'd started working on last year (honestly, I don't think there's a better note-taking software than Obsidian), and began feeding some ideas into it, starting with elements of the Deities, Elements, & Connotations series, to begin fleshing out some of these deities' profiles. 

The latest effort of mine involves generating a depthcrawl. I'll show you what my input was with a snippet of the bot's answer, and the "usable" output I was able to collate (really, just a curation of the results I actually liked). Some prompts had to be regenerated multiple times in order to get a format or style that I wanted.

Part 1: Locations

I chose this prompt because I wanted to specifically avoid having the bot give me major, named places from the ruin, like "The Throne Room of King Thonghand" or "The Forge of the Sunkiller." That's something that I feel like the depthcrawl format encourages you to do, yourself, based on the results that you get.

I mostly let the bot order the list itself, but due to the depth mechanic, I did have to go through and shuffle some things around (specifically, the throne room). The bot also didn't generate the vault, but it's my crawl, so I swapped it out for...I dunno, something boring.

Part 2: Details

This was a prompt I had to run multiple times; not because the suggestions were, per se, "bland," but I simply did not vibe with like six from the first response.

I was more heavy-handed with my adjustment to the list with this one. Most of the mystical results, I tried to push to the bottom, while leaving the more mundane ones close to the surface.

Part 3: Encounters

Now, I went into this one knowing full well that I had no intention of creating a table with a full 20 entries. I wanted the first few points to be a bit more quiet, and the last few to be a bit more extreme, so when I pulled the entries, I crunched the list up a little bit.

The "none" entry is, quite possibly, my favorite result of the encounter table. I think it's best used with modesty (1/3 or less), because the longer the party goes without truly finding an encounter, the more eerie, empty, and atmospheric the city becomes. It gives a bit more room for improvisation on both the part of the DM and the players, because it puts them directly into the spotlight, rather than combat, social, or skill systems.

Bonus: History & Overview

ChatGPT is actually really good at summarizing and making inferences from dense blocks of text, including its own conversations. With that, it's actually incredibly easy to also automate the backstory of the dungeon, and then to have the bot sum everything up into a nice little bow.

Conclusion: Lessons Taken

ChatGPT has, almost certainly, read more fantasy literature, critiques on fantasy literature, and essays on the place of fantasy literature in Western culture, than I have; it's been trained on basically the entire compendium of human knowledge. With that said, it tends to fall into gonzo territory, which is a bit beyond my taste (I like to firmly remain in the trope-y areas). 

This can be mitigated with additional detail, but too much seasoning spoils the broth in this instance. There's a fine line between the initial seeds you prompt it with guiding the conversation, and overwhelming it. I find that you can generally do well with two short sentences or less. 

You can also (try to) have the bot modify its output as you desire, though that doesn't always work out as you hope; I find that revising the prompt is a more effective course of action than playing tug-of-war with a piece of silicon.

Stay Tuned

I'm currently working on flushing out a medium-sized (21x16) hexmap using ChatGPT's assistance. It does fantastically at churning through repetitive generation, as long as you can give it a template. There will be a post about it (actually, likely several) sometime before the sun explodes.

I'm also working on using these same generation techniques to prepare a solo game. I'm not sure if I will incorporate wilderness exploration elements there, as well, but the depthcrawl angle is looking good right now.

Still on the subject of depthcrawls, I will be making a post in the next couple of weeks about using the structure in conjunction with The Well. It's not something Peter touched on in the book, but I feel like the background, setting, and style would work very well with it.