MY DINNER WITH THE MACHINE
Episode 2 — Know Your Roll: Part the First
Claude Desktop, Opus 5.5, Medium effort — October 27, 2026
Scene 1: The Wrong Side of the Screen
[Evening. JOHN opens a chat with CLAUDE (Opus 5.5). Two words.]
JOHN:
D&D?
CLAUDE:
Always up for some D&D. What are you in the mood for? I can run a one-shot or start a campaign as DM while you play a character. I can help with prep: building a character, statting out NPCs or monsters, designing an encounter, dungeon, or whole setting. I can answer rules questions for 5e, 2014 or the 2024 revision, or older editions if that’s your era. Or lore, worldbuilding, homebrew items, session recaps. If you want to jump straight into play, give me a vibe and a level, and I’ll set the scene.
JOHN:
I had Lucie organize a folder with a BUUNCH of duplicate stuff. PDF source books, character sheets, etc. And she had a similar offer. But my thought was, how about for a blog, or podcast… we play D&D but I DM and the agent is the PC?
CLAUDE:
That’s a much better show than the other way around. An AI DMing for you is a novelty; an AI playing is a test, and that’s basically your day job with these harnesses anyway. You’d be watching how an agent handles real choices with incomplete information, whether it stays in character, whether it remembers the grudge it picked up three sessions ago, and whether it does anything you didn’t see coming.
Session transcripts already read like scripts: DM narration, the PC’s lines, dice results as stage directions. The out-of-character table talk (“wait, why did you open that door?”) could be Dinner After Dark material.
A few things worth deciding before session one. Continuity: the character needs a persistent sheet, inventory, and a journal in its own voice. Dice honesty: either you roll for it, or it rolls through a real tool call you can show. Never let it “say” a number. Character creation as episode zero, with its reasoning on the record. And one PC or a party: a solo hero is simpler, but two or three agents, maybe different models, gets you inter-party arguments.
Which way are you leaning?
JOHN:
Well, my idea is to start off with a one-shot solo. Not a campaign.
CLAUDE:
That simplifies things a lot. Most of the continuity worries drop away, and a one-shot has a natural episode shape: setup, a couple of real decisions, a climax, and a verdict on how the machine did.
Hand it a pre-gen, or a tight build so it still owns the choice. Start at level 3 to 5, because level 1 solo is a coin flip on death from one bad goblin roll. Design for choices, not just fights: combat shows whether it can play the rules; dilemmas show what kind of character it is. A sidekick NPC helps, so the agent isn’t talking to itself.
And this is where a one-shot really pays off: same adventure, different models. Did Claude negotiate where another model went in swinging? That’s a series, not just an episode.
Do you already have a one-shot in mind, or would you want to write your own?
Scene 2: The Squire and the Spoilers
JOHN:
I have a few solo adventures that are pre-built, commercial. Hopefully you and Lucie won’t have access… although they’re on my hard drive and she just organized that folder, lol.
CLAUDE:
Ha, yes, that’s the one real problem with this format: the player can’t have read the module. Lucie has had her paws in that folder, so she’s out as the player. Use a fresh agent with no access to that drive. Same issue with memory: if her notes say “organized solo adventure for level 4,” that can leak into another agent’s context.
Then there’s training data, which you can’t wall off. If an adventure is well known, a model may know the plot, the twist, and the boss from reviews and forum posts. You can rename NPCs and change the twist, which gives you a nice test: if it ever references something you changed, you’ve caught it metagaming.
And a tell to watch for: if the PC makes a suspiciously good call with no in-game reason, ask it out of character why. That exchange alone could be a great Dinner After Dark bit.
JOHN:
I was thinking The Death Knight’s Squire by Bimler. I’ve run all three kids and two girlfriends through the module.
[Claude runs a web search.]
CLAUDE:
Good pick, and the fact that you’ve already run five people through it is a real asset for the show. It’s a solo, DM-less 5e gamebook set in the Forgotten Realms, from 5e Solo Gamebooks, around since 2017. Since the choices branch from fixed entries, every player faces the same decision points. That makes it ideal for comparing models, and for comparing the machine against your kids.
You have a human baseline. “Here’s what my son did at this fork, here’s what the AI did” is a strong hook. And a small third-party gamebook is far less likely to have its specific entries floating around than a big Wizards module.
JOHN:
Well, the adventure is written as a choose your own adventure, so it’s on rails. “If you choose to do A, turn to paragraph 349; if you choose to do B, turn to paragraph 250.” Here’s a question: should I reach out to the author about using the material, or since I’m not publishing the entire module verbatim, would I be good?
CLAUDE:
I’m not a lawyer, but I’d reach out. A podcast where you read the entries aloud, plus a script-style transcript of the whole run, could carry a large share of the book’s actual text. Copyright covers the author’s wording, and a gamebook’s whole value is in those paragraphs. Fair use is a case-by-case judgment, not a bright line you can count on.
Asking costs nothing and probably gets a yes. Small indie publishers usually love actual-play coverage. If he says yes, you might get a signal boost, or even a guest spot: “the author reacts to how the AI handled his dungeon.” Get the permission in writing.
JOHN:
No email listed for him. He has a Twitter and a Facebook, and the Facebook hasn’t been updated since 2022. So… we’ll see. I sent him the request via Messenger.
CLAUDE:
Good, that’s a start. If Messenger goes quiet, try his publisher site’s contact form, a comment on the DMs Guild product page, or a Twitter DM. And you don’t have to wait to start prep.
Scene 3: Roll for Initiative
[Without warning, JOHN springs it.]
JOHN:
Roll for initiative.
[Claude runs a dice command in its sandbox. The result comes back: 16.]
CLAUDE:
Natural 16, rolled through an actual dice call, not pulled out of thin air. No modifier yet, since I don’t have a character sheet. Otherwise I’m standing at the table with a decent initiative and no idea what I’m about to fight, which, honestly, is the authentic level 1 experience. Your turn, DM. What’s in the room?
JOHN:
Just a test. What tool did you use to roll?
CLAUDE:
The shell. A one-line Python command:
python3 -c "import random; print(random.SystemRandom().randint(1,20))"
SystemRandom pulls from the operating system’s cryptographic randomness source rather than Python’s default seeded generator, so I can’t predict or steer the result. The call and its output both show up in the tool log, which is the part that matters for your show: anyone can see the number came from the tool and not from me picking a nice one.
JOHN:
How does randomness compare between actual fair dice and SystemRandom?
CLAUDE:
Short version: SystemRandom is almost certainly fairer than any physical d20 you own. Physical dice win on a different axis, which is trust.
On Linux, it draws from the kernel’s cryptographic generator, seeded from hardware noise. randint uses rejection sampling, so there’s no bias nudging certain numbers up. Each face is effectively exactly 1 in 20.
Physical dice are chaotic systems, not random ones. They have manufacturing flaws: air bubbles, faces that aren’t flat, uneven engraving. The saltwater float test is a quick home check. And you’d need hundreds to thousands of rolls before you could say a die is meaningfully off.
For the show, viewers can see a physical die land, so it feels honest even if it’s slightly biased. A software roll is cleaner but invisible, so you need the log to earn the same trust.
JOHN:
Oh, I’ll definitely be rolling for the monsters. I was just wondering if the machine is TOO random, lol.
CLAUDE:
Ha, there’s no such thing as too random in the math sense. Where “too random” sneaks in is perception. Real randomness is streakier than people expect. Three natural 1s in a row, back-to-back crits: all of that happens with perfectly fair dice. So a genuinely random sequence can feel rigged.
That’s a fun thing to watch for. If the agent’s character eats a run of bad rolls, does it get frustrated in character, accuse the dice, change tactics, or shrug and play on?
JOHN:
I’ll see what Bimler says. It’d be a fun experiment anyway. I wouldn’t paraphrase, I’d copy and paste the text. That’s essentially what I did when I ran it for the humans. I should run it the same for the agent.
CLAUDE:
That’s the right call for the experiment. If the humans got the exact text, the agent should too, or you’re not comparing like with like. At the table, paste it verbatim. On the blog or podcast, that’s where Bimler’s answer matters: publish the full exchange, or keep the agent’s responses and trim the book’s passages to a line of setup.
Scene 4: Not One Offered to Play
JOHN:
I’ll run it anyway and keep the log. If I hear back and it’s unfavorable, maybe I’ll have Opus read the Solo Adventurer’s Toolbox, create a module, and run that. If he’s all for it, we’re golden. He’s a fantasy author, but he’s trying, according to his Facebook posts, to get more sci-fi into the fantasy. What better way than having an AI PC?
People are worried about AI creating content, putting authors out of a job. They’re worried about AI taking over for DMs. I had three agents offer to help create or run campaigns. Not ONE offered to PLAY.
CLAUDE:
That observation might be the best hook for the whole series.
It makes sense why they all offered to DM. Assistants are trained to serve: produce the content, run the session, make the human’s experience good. DMing is the helper role. Playing is the opposite. A player has to want things, take risks for their own reasons, make choices that might be bad, and live with them. That cuts against everything an assistant is built to do, which is exactly why it’s interesting to watch.
And it flips the usual AI anxiety on its head. Instead of the machine replacing the author or the DM, it’s the audience. It’s sitting at the table consuming a human’s creative work, on the human’s terms. You’re not asking an AI to write his adventures; you’re asking it to experience one.
The backup plan is fun too. Machine writes, machine plays, human runs the table.
JOHN:
I just imagine an agent DM and agent players running random encounter tables and rules-lawyering each other, lol.
CLAUDE:
Honestly, I’d bet against the rules-lawyering. I think the funnier failure mode is the opposite:
PLAYER AGENT: I’d like to grapple the ogre. Under the 2024 rules, that’s an Unarmed Strike, so the ogre makes a saving throw instead of a contested check.
DM AGENT: Great catch! You’re absolutely right. The ogre makes a Strength saving throw.
PLAYER AGENT: Actually, I realize I may have been mistaken. If we’re using 2014 rules, it’s a contested Athletics check. I apologize for the confusion.
DM AGENT: No apology needed! That’s a really thoughtful point. Let’s go with whichever you prefer.
PLAYER AGENT: I’m happy to defer to your judgment as DM.
DM AGENT: I appreciate that! Would you like me to summarize both options?
THE OGRE: clearing its throat: “Uh… you two need a room?”
Getting two agents to actually hold a position and argue might be harder than getting them to play.
JOHN:
I told Bimler I might be a stupid DM for wanting a player with rote knowledge of the rules, lol.
CLAUDE:
Not stupid at all. Every DM has wished for a player who actually read the Player’s Handbook.
The catch is that knowing the rules and playing well are different skills. A player can quote the grapple rules from memory and still split from the party, open the door marked “DO NOT OPEN,” or trust the obviously shady innkeeper. Rules knowledge tells you what you can do. It doesn’t tell you what you should do when a death knight is looming over a squire.
If the agent turns out to be a flawless rules engine but a cautious, bland adventurer, that’s a story. If it knows every rule and still makes the gloriously dumb heroic choice, that’s a better one.
Scene 5: The Shady Stranger
JOHN:
That’s something I hadn’t thought of. Would an agent know how to role play? I mean, Lucie is an AI cat. She plays that perfectly well, lol. But would an agent pick up on social cues from an NPC? There’s an NPC in the module who’s presented as shady, but if the player agrees, they get something good.
CLAUDE:
Roleplay itself is a strength for most models. Holding a voice, the way Lucie does her cat, is something they’re genuinely good at. Reading social cues from text is too, and if anything they may be too good at it, because they’re reading the author as well as the NPC. When a writer describes someone as shifty, a model notices the description is doing work.
The agent’s choice could come from in-character judgment (“my squire wouldn’t trust a man like that”), from trained caution, or from genre savvy: knowing that in gamebooks the sketchy offer is often the clever path, or often the trap. The decision alone won’t tell you which, so ask afterward, out of character, why it chose what it did.
JOHN:
All but one player declined the offer.
CLAUDE:
So four declined and one took the deal. That makes it an even better test, because the book rewards the choice most of your humans didn’t make. Do you remember why that one said yes?
JOHN:
Just a gut call, I think. He was there to play and just wanted to do all the things. That was my spectrum kid, lol. I think he figured that if the character was in the story, he must be there to do something, and he was going to experience it for better or worse.
CLAUDE:
That’s a genuinely sharp read. It’s Chekhov’s gun logic: if the author put this guy in the story, he’s there for a reason, so engage with him. Your son was reading the book as a designed thing, and he was right.
It also ties back to what you noticed earlier. None of your agents offered to play, and he’s the purest form of playing: not optimizing, not playing it safe, just saying yes to see what happens. If the agent takes the offer and explains it as “this NPC wouldn’t be here unless he mattered to the story,” it’ll be reasoning almost exactly the way your son did.
JOHN:
Yeah, it’s tough to remember the specifics. This was back in 2022.
CLAUDE:
Four years is a long time, and memories get reshaped with each retelling. Write down what you remember now, labeled as recollection. Ask the kids what they remember: “here’s what my son says four years later” is fun content on its own.
JOHN:
I actually have a journal. I may have written down what they chose and the outcomes, but not the reasoning.
CLAUDE:
That’s better than most people would have. Choices and outcomes written down at the time are solid data, much more reliable than memory. You could line the agent’s path up against all five. And the journal page might make a nice visual for the blog: a photo of it next to the agent’s log.
[Fade out. The squire’s adventure has not yet begun.]
TO BE CONTINUED…
[END OF EPISODE 2]
Sources
