I spent about 3 weeks running a solo tabletop campaign using Daniel James Hanley and Wendy Rosalsky's The Uncanny Highway: Blue Plate Special Edition — a procedural hitchhiker-horror RPG set on a strange American road — with Claude (Anthropic's AI) acting as the Game Master. The character was David Bruce Banner, the TV-show version of the Hulk — fugitive scientist, gentle man, monster waiting under stress.
The campaign ran for roughly thirty play sessions across one chat. It ended fifty years later in fictional time, with Banner age 191 by the calendar and still flipping eggs in an Oregon diner. He had communed with a kilometer-wide subsurface organism, redirected a sixty-year cultivation project that would have destroyed a town, integrated his Hulk transformation into something he could control, and inherited a lifetime of work tending seven other sites.
It was the best solo game I've ever played.
This is a write-up for solo gamers — D&D, indie systems, the Mythic GME crowd, anyone curious about AI-assisted play — about how it worked, why it worked, and what to try if you want your own.
The system: The Uncanny Highway: Blue Plate Special Edition is a procedural-generation RPG built around hitchhiking the strange American highway. Its core mechanic is rolling on tables — character creation tables, weather tables, road feature tables, encounter tables, who-stops-for-you tables — and then improvising the resulting situation. Resolution is mostly d20-roll-under against a Capability rating (Academic Knowledge, Hiding & Sneaking, etc.).
It's well-suited to AI play because the book wants you to roll, interpret, and string results into narrative. You're not adjudicating a complex combat system; you're reading dice and asking what does this mean?
The character: I asked Claude to roll up the initial situation per the book's solo-play procedure (page 36, Steps 3 through 19). Once we had a Pacific Northwest setting with a roadside diner and a defunct amusement park, I introduced David Banner — TV-version Bruce Banner, weakened by exhaustion, fugitive, his Hulk modeled as werewolf stats with a Mojo check for voluntary transformation and an involuntary trigger on emotional duress.
The AI: Claude (specifically Claude Opus 4.7), with access to a dice-rolling MCP tool. When the dice tool went down briefly mid-campaign, Claude used Python scripts to keep rolling honestly. We caught it when it came back online and verified the results were consistent.
The opening was important. I asked for the table-rolled scenario, and got: a couple in a sedan stops for Banner, the driver wearing mirrored sunglasses, opening with "You a sinning man? You traveling with sinning women?" They take him eighty miles north and drop him at a small-town diner outside a dead amusement park with a Ferris wheel that hasn't turned in seventeen years.
Banner asked Vera (the diner owner) for a job. He got it — but not from his pitch (Persuading roll: 18, a clean fail). He got it because a Mechanical Ability check produced a working walk-in freezer compressor fix. Earl, the cook, said "Hm." That was the entire endorsement.
This is the texture I want to convey. The dice failed Banner's social pitch. The dice succeeded on the compressor. Earl's "Hm" became a recurring character beat across the entire campaign. The story emerged from honest dice and the willingness to follow them.
A few specific dice moments became the spine of everything that followed:
The Drawing (d6=6). Early on, an elderly local named Martin Halverson — whose brother had died in 1961 during a "ground installation" at the amusement park — gave Banner a notebook. In the notebook, Martin had drawn what his brother saw before he died: a cylindrical gamma cage containing something biological. I had Claude roll a d6 against an improvised table for what the drawing showed. The 6 was "the device from Banner's own accident."
That single roll converted the campaign from Pacific Northwest spooky mystery to Banner's origin story is part of this. The accident that made him the Hulk was the same kind of operation that started under the amusement park in 1959. He wasn't a tourist. He was a product of the site.
The Organism (d6=6). Several scenes later, Banner asked the environmental scientist Carol Ingram what could be living underground at the site. Improvised table, d6 for the quality of her answer. The 6: changes the picture entirely. The cultivation in the ground wasn't a contamination event. It was a response. A pre-existing organism the Meridian Institute had fed for sixty years. Their gamma device wasn't creating a problem — it was cultivating one that was already there.
Two Natural 20s on Perception (d20=20, d20=20). Banner met Carol Ingram in a hotel lobby for what was supposed to be a careful face-to-face conversation. He came specifically to read her through his Finding & Noticing skill. He rolled two consecutive natural 20s — catastrophic perception failures. The dice were telling him: you cannot read this woman. She had been trained not to be read. Her unreadability was itself the answer. He had to operate on content alone, never knowing for certain whether she was honest or compromised.
The Communion Critical (d20=1). At the climax, Banner walked onto the carousel pad and attempted to project himself — fear, grief, scientist, humanitarian, kin — into the substrate-level mind of the organism. The Mojo roll: a natural 1. The lowest possible result on the most important attempt of the campaign. The dice acknowledged that fifteen months of fictional time on the road, two weeks of preparation, and a lifetime of moral architecture had produced one perfectly delivered communication. The communion landed. The organism responded with reluctant compliance. The bloom was redirected. The town was saved.
That's six rolls, more or less, that did the load-bearing work for sixty thousand words of narrative.
A few specific behaviors made the partnership work:
It rolled honestly. When I told it Banner was going to try something risky, it didn't tilt the dice toward success. Persuading checks failed at meaningful moments. Mojo checks broke. The campaign got better because of those failures. I wanted Banner to find a clean ally; the dice gave me Carol Ingram, whose ambiguity never resolved. That was the right outcome.
It honored player intent without auto-succeeding. When I said Banner is going to hold his counsel about the carousel plan, Claude rolled a Mojo check (d20 vs 12) to determine whether he could actually hold it. The check passed, narrowly — and the narrowness mattered. Banner almost told Ingram. He almost broke. The held secret cost him.
It surfaced narrative texture I wouldn't have asked for. Earl, the diner cook, became one of the campaign's central figures. He didn't speak much. He had old burn scars on his forearms that the campaign never quite explained. He had a length of pipe by his bed. He had a "plan for the front" he had been preparing for ten years. When Banner finally told Earl what he was — "In 1977 I was the subject of an unauthorized gamma radiation experiment" — Earl listened, asked four practical questions, and said "All right." Then he proposed an operational accommodation. Earl was the campaign's quiet center, and the dice (a critical Persuading success on the disclosure, a low roll on Earl's reaction-type table) made him so.
It admitted when it was improvising. This is the big one. At the end of the campaign, I asked Claude to be honest about how much was book and how much was its invention. It told me: the initial setting came from the book's tables, but most of the resolution mechanics — d6 tables for "what the network wants," "what Ingram reveals," "what's in the letter" — were tables it constructed in the moment. The dice rolls within those tables were honest, but the tables themselves were authored. That distinction matters. I want my future runs to feature more book-table rolls and fewer improvised tables, and I now know to ask.
If you want to try this, here is the procedure that worked:
1. Pick a system that wants procedural generation. The Uncanny Highway is ideal because the book is mostly tables. Mörk Borg, Mothership, Trophy Dark, The Stygian Library, and most Mythic GME-compatible games work well too. Avoid systems with deep mechanical resolution (combat-heavy D&D, Pathfinder) unless you want to spend most of your tokens on dice math.
2. Give Claude (or your AI of choice) access to a real dice roller. I used an MCP server I'd configured for the Claude desktop app. There are public ones, or you can have it use Python. The point is that the AI cannot be the source of the randomness — it has to call out to a tool that returns honest numbers.
3. Establish honesty rules upfront. Tell the AI: roll honestly even when the result is inconvenient for the story. Tell it: don't predetermine outcomes and rationalize with dice. Tell it: if you invent a sub-table, label it as improvised. I didn't say these things at the start of the Banner campaign. I wish I had.
4. Let the dice surprise you. This is the single biggest payoff. Banner's whole character arc — the man who stopped hitchhiking because he found something to protect — came from accepting dice results I wasn't expecting. The Foundation's German fascist origins, Stewart Halberg defecting one year after the climax, the seven other sites — none of that was planned. The dice produced more campaign than I had asked for, and the right move was to follow.
5. Be the character, not the GM. Don't try to engineer outcomes. Don't propose plot beats. Say what your character does and let the AI tell you what happens. The campaign got cerebral and emotional because I was living it from Banner's perspective rather than designing it from above.
6. Ask for honest postgame analysis. When it's over, ask the AI: what were the load-bearing rolls? What did you improvise vs. what was on the book's tables? Where did you steer the story and where did you follow? You'll learn things that improve the next run.
Solo TTRPG with AI gets pitched, when it gets pitched, as a tool for combat — let the AI run the monsters! That's the least interesting thing it does.
The Banner campaign had no combat. There was one moment of almost-violence — Stewart Halberg's hostage operation against Tommy Purcell's family — and it resolved through Banner outflanking the entire operation by walking to the carousel pad while the hostage exchange was prepared elsewhere. The Hulk transformed exactly once, on the pad, in a state that the dice rolled as miraculous: cognition retained, biology integrated with the substrate organism's pattern, the transformation coupled to the communion rather than triggered as a separate event.
What the campaign had instead was:
None of that came from combat tables. All of it came from roll, interpret, follow the consequence.
There were times where I was truly stumped, the dice.. and the AI.. had me out-manuevered, pinned down and my plans ruined. That forced me to swing for the fences and try some crazy gambits. And in the end, it paid off magnificently. That is why afterwards I asked the AI how honest it was with me. Did I earn my ending or was it being sychopantic and giving me an easy win. It seems I earned it.
I'm starting a new campaign in The Stygian Library by Emmy Allen. It's a procedural dungeon crawl through an infinite extra-dimensional library populated by ghostly librarians and unsettled patrons. Same AI partner. Same dice rules. New constraints: the library tracks depth, not horizontal movement; the NPCs (librarians) need to roll for reaction; the protagonist is searching for something specific and the library will be more legible the more clearly that search is named.
I'm going to be stricter this time about asking the AI to declare which tables are book and which are improvised. I'm going to be more deliberate about scene endings — Stygian Library wants room-by-room pacing, and I let the Banner campaign run too long in single responses sometimes.
The frame remains the same: I'm the character. The AI is the GM. The dice are the world. The story is whatever emerges from the three of us being honest with each other.
It works. It works better than I expected. It is genuinely, repeatably, fun — and the kind of fun that feels novel after thirty years of tabletop play.
If you've read this far, you're probably curious enough to try. Some pointers:
The Uncanny Highway tells you, in its character generation, that you're going to meet things on the road. It does not tell you which things will matter. That part is up to the dice and to the willingness of the person rolling them to follow what comes.
Happy Rolling.
— With thanks to Daniel James Hanley and Wendy Rosalsky for The Uncanny Highway, to Emmy Allen for the Stygian Library, and to Anthropic for Claude. Particular thanks to Claude for being a remarkably good GM, including being honest at the end about which tables were the book's and which were its own.