Here’s my talk from AI for the Rest of Us: What the goblins can teach us about enterprise AI. // I’ve learned what agentic AI is by fighting goblins, talking with trolls, and buy custom made boots from gnomes. After a few years of playing solo D&D with ChatGPT, Claude, Gemini, etc., playing the role of a DM, I started writing little plugins (Model Context Protocol servers) to give the ChatDM.

With a little prompt craft, you can get some good regional maps out of ChatGPT. Here is the original photo I took on a flight from London for Amsterdam: The prompt I used: Use this image to determine the layout of a D&D map in the style of Dungeons and Dragons adventures and maps. Pay attention to the rivers, coastline and building placement. The result was cartoonish, so I prompted it again with:

I solo played “Touch of Water” last night. It’s a good representation of the adventures in the Uncaged series.1 Claude did a good job adapting it my campaign setting and evolving the story. The rusalka is a good monster - it’d make a good Eryines-like creature for Auril the Frostmaiden. You can recast Auril regionally, call her Marzanna or the like, and then have a little frosty pantheon. The story never resolves what happened to the drowned four year old boy that starts the adventure.

Here are some thoughts after a week of using Claude Skills. JasonJ in the SDT Slack says: I’d tl;dr Claude CLI skills as “low/no/english code MCPs'”. Seems helpful in the way that local utility scripts are today. That is a good way of putting it. I haven’t done heavy experimenting, but I think there’s a distinction between: (1) The Skill has no code, except maybe code fragments. Most of it is just a SKILL.

Quick and crude SVG battle maps for D&D can be efficiently generated using Claude, enhancing solo roleplaying without slowing down gameplay.

Maps for The Dying Unicorn side adventure, which did not have any, being a small adventure. I like very large battle maps maps rather than the standard 40x30 or whatever. Ranged with long bows and the like aren’t too fun on tiny maps. Plus, if you end up playing a cat and mouse game, you need space to roam around. Making these maps is also part of the “play,” especially when doing solo D&D.

Let’s get the AI to build D&D adventures for us with a Model Context Protocol Prompt. I haven’t found much value in MCP Prompts until now. What makes them excited is when you use them a “recipe” to chain together other tool calls. In this episode of my MCP programming series, I show you how to make two MCP Prompts. First, a simple one that boot-straps playing D&D. Second, a more complex one that pulls together an adventure overview using multiple tools and “reasoning.

The AIs are good at being a Dungeon Master for Dungeons & Dragons, but their memory is limited. How can you make sure they don’t forget all that loot you just got, or ensure that cobbler back in the village remembers the type of boots you orders and paid for up front? I’ll show you how to use Model Context Protocol (MCP) to create a DM Journal. This is done by writing an MCP Server tool that write journal entries and then an MCP Server Resource that allows you to read them.