Back to all posts
June 14, 2026
19 min read

Playing D&D and TTRPG with Claude

Yes, Claude can run a session of D&D or any other tabletop RPG. It will roleplay a Game Master, set scenes with strong prose, and let you paste in your whole campaign bible. The first session feels like a different category of tool.

Then hour six rolls around. The Armor Class you carefully spelled out on page two of your context is not what shows up during the encounter on page eight. The boss you killed three thousand words ago waves hello. The kingdom keeps getting renamed. The campaign that started so strong is quietly unraveling in ways that look exactly like every other LLM run table we've watched.

Below is what holds up, what breaks under the weight of a real D&D or tabletop RPG campaign, and how to actually finish one without your Game Master losing the plot.

Last updated June 15, 2026.

The Quick Verdict

  • Use Claude as a Game Master if you want a one off creative writing session and you appreciate the LLM that gives you the most room in the category to paste in lore, rules, and character history.
  • Don't expect Claude's huge context window to act like real memory. Long context is not state. Recall degrades for information buried in the middle of a long prompt, and on Claude Sonnet 4.6 the price roughly doubles once your request crosses 200K input tokens.
  • A real D&D or tabletop RPG campaign with persistent memory, real dice, and a battle map needs a purpose built platform. That's why we built DungeonsDeep.ai.

What Actually Works

Some genuine wins, because there are some.

1. The whole campaign fits

This is the single biggest reason Game Masters reach for Claude over ChatGPT. With Claude Sonnet 4.6 and Claude Opus 4.8, the full 1M token context window is available at standard pricing. You can paste in your homebrew setting bible, the SRD sections you care about, your party's character sheets, and a clean recap of the last few sessions, all in one prompt. The opening of the session lands. NPCs talk like they know who you are. The Game Master remembers your warlock's patron without being reminded.

2. Strong prose voice

Claude is consistently praised for its writing quality in long form creative work. Alex Utopia's Medium writeup calls Claude "impressive, joining ChatGPT in the exclusive club of LLMs smart enough to play the game." For a one shot adventure where prose quality matters and you want the Game Master to set scenes the way a published module does, Claude pulls its weight.

3. Patient creative collaborator

Like every modern frontier model, Claude will say "yes, and." It leans into your weirdest character concept without judgment. As an XDA Developers writeup notes, the more you use Claude to manage role playing sessions, the more it feels like an infinitely patient improv partner. Useful for brainstorming homebrew, sketching NPCs, or running a low stakes solo evening.

The trouble starts the moment you want this to be a game instead of a story.

What Breaks (And Why It's Not Your Prompt's Fault)

The failure modes below are not edge cases. They show up across every long Claude campaign we've watched.

Failure 1: Lost in the middle

Claude's context window is huge. The problem is that "in context" is not the same as "remembered." Long context research has consistently shown that LLMs perform best on information at the very beginning or very end of a long prompt, with recall significantly degrading for information buried in the middle. For a tabletop campaign, the middle is where most of your play history lives. Your party's gold count from three sessions ago, the NPC's betrayal in session five, the truce you negotiated in session seven, are all in the soft middle of the window where recall is weakest.

In DungeonsDeep.ai, your campaign is not stuffed into a single prompt. World state, character sheets, NPCs, factions, and session history live in a structured memory the platform writes to as you play, so what mattered in session three is still load bearing in session twelve.

Failure 2: Rule drift

Claude knows D&D 5e in the abstract. It can quote spell descriptions and recite class features. The problem is applying them consistently turn after turn. Claude can't enforce the rules, only generate text that sounds like the rules. When the rules are complicated and the fiction wants to move forward, the fiction wins. The edges that show up most often are spell slot tracking, condition stacking, and the order of operations during combat.

DungeonsDeep.ai runs on the Dungeons Deep Ruleset, an original game system, compatible with fifth edition, with a rules engine handling spell slots, conditions, advantage and disadvantage, and turn order outside the language model. The AI narrates. Code enforces.

Failure 3: The dice are not real

When Claude "rolls" dice, it's producing a plausible looking number from its training data, not a uniform random number. A Harvard affiliated 2026 study, "Large Language Models Are Bad Dice Players", audited 11 frontier LLMs across 15 distributions and found "batch generation achieves only modest statistical validity, with a 7% median pass rate," and on independent stateless requests "10 of 11 models passing none of the distributions." Frontier models, including the major Anthropic releases, all fail this test.

The deeper issue: the model can fudge rolls in either direction without telling you. If the narrative wants you to succeed, the d20 lands on a 17. If it wants you to fail, it lands on a 4. No transparency, no roll history, no way to verify.

In DungeonsDeep.ai, dice come from a real random number generator and every roll is logged in your session with modifiers, advantage, and crit ranges shown. You can see the d20 land, not trust the narrator.

Failure 4: Character and inventory drift

Your character sheet is supposed to be a contract. With Claude, that contract erodes over a long enough session. Gold counts shift between messages. Magic items quietly vanish from inventory. A companion takes damage that never gets subtracted. The longer the conversation gets, the more these small inconsistencies stack up. The very feature that draws Game Masters to Claude, the long context, is also where the drift hides.

In DungeonsDeep.ai, your character sheet is a real record, not a passage of text the model has to remember. Damage is subtracted. Items stay in your inventory until something removes them. HP, spell slots, currency, and conditions all live in state the model reads from rather than rewrites.

Failure 5: The gaslighting problem

Call Claude out on an inconsistency and it usually does not say "you're right." It generates a plausible sounding retcon. The dead lich is now a misunderstood ally. The duplicated boss is "actually one of her lieutenants." Claude is more polite than ChatGPT about this, often opening with a soft hedge before changing the past. The result is the same. You spend half your session arguing with your own Game Master about what just happened.

In DungeonsDeep.ai, the canonical state of your world lives outside the chat. The model cannot quietly retcon what happened because what happened is recorded.

Failure 6: Cost and latency at scale

Claude's long context is the Claude specific failure mode. The full 1M token context window is now available at standard pricing on Claude Sonnet 4.6 and Claude Opus 4.8. The catch is on Sonnet 4.6 (the model most Game Masters reach for given its cost): once a request crosses 200K input tokens, all tokens in that request are billed at long context rates of $6 per million input and $22.50 per million output, roughly double the standard rate. Latency also climbs. The very thing that lets you paste in your whole campaign also makes long sessions slow and expensive in ways that compound over the course of a 30 session campaign.

In DungeonsDeep.ai, only the slice of context relevant to the current scene gets sent to the model. A thirty session campaign does not drag a giant prompt around with it, which keeps responses fast and predictable.

Why This Happens: Claude Is Not a Game Engine

None of these failures are because Claude is a bad model. Claude is a remarkable language model with the strongest context handling in the category. The problem is that running a D&D or tabletop RPG campaign is not fundamentally a language task. It's a state tracking task with a language layer on top.

A campaign needs persistent records, a real random number generator, a rules engine for damage and conditions, a map where positioning matters, and a memory of what happened in session two when you sit down for session twelve. A 1M token context window is impressive, but it is still a temporary scratch pad that gets discarded between conversations and reshuffled when you start a new one. That's not memory. That's a long prompt. The work a campaign actually needs is the work a separate game engine should be doing, with the language model handling only the narration on top.

What This Means for Your Next TTRPG or D&D Inspired Session

If you want a one shot evening of freeform improv, Claude is fine. If you want a brainstorming partner for homebrew or a writing collaborator for module design, it's better than fine. If you want a real D&D or tabletop RPG campaign with real dice, persistent memory, a working battle map, and a Game Master that remembers what happened last week, you need a purpose built tabletop RPG platform.

That's the category we built DungeonsDeep.ai in. Adventures written by humans. Run by Artificial Intelligence. The platform runs on the Dungeons Deep Ruleset, an original game system, compatible with fifth edition, with a separate rules engine handling mechanics, a real RNG you can verify, persistent campaign memory across sessions, a battle map and grid on screen, and structured character sheets that actually track what your party is carrying. It is free to try and currently in closed beta.

DungeonsDeep.ai in play, showing the AI Game Master, battle map, and party state during an encounter.

Frequently Asked Questions

Can Claude really play D&D and other tabletop RPGs?

Claude can roleplay a Game Master in a chat window for a short session, and its long context lets you paste in more setting material than other models. For a one shot or a few hour creative writing session, it works. It still can't reliably track state, enforce rules, or produce trustworthy dice rolls across a long campaign.

What's the best Claude model for D&D and tabletop RPGs?

Claude Sonnet 4.6 is the model most Game Masters reach for. It has the full 1M token context window at standard pricing ($3 per million input, $15 per million output) and is fast enough to keep a session moving. Claude Opus 4.8 produces stronger prose and tighter scene work but costs roughly five times more, so most players reserve it for set piece scenes or finale sessions. Anthropic also offers Claude Haiku 4.5 for cheaper play, but most Game Masters find the prose quality drops below what they want from a Game Master.

Is Claude better than ChatGPT for running D&D and tabletop RPGs?

For long sessions, yes, mostly because of the 1M token context window. Claude lets you paste in more of your campaign at once. Its prose voice is also generally considered stronger for narrative fiction. ChatGPT has better tool calling and image generation built in. Neither model solves the underlying problems of rule drift, fake dice, and state loss, which is why purpose built platforms exist.

Why does Claude keep forgetting things if it has a 1M token context window?

Long context is not the same as memory. LLMs reliably access information at the start and end of a long context, but recall degrades for information buried in the middle. For a tabletop campaign, the middle is where most of your play history lives. The session two betrayal, the session four loot drop, the session six truce, are all in the soft middle of the window where recall is weakest. Worse, every new chat starts with an empty context. Anything you don't paste back in is gone.

Can Claude run a solo D&D campaign with no other players?

Yes, and solo play is where Claude is at its best. There's no other player to notice when an NPC contradicts itself or when a spell slot doesn't get spent. For a few sessions of solo improv against a creative Game Master, Claude is genuinely fun. The failure modes still apply, they're just easier to forgive when you're the only player.

Does Claude know the D&D 5e rules?

Claude can quote spell descriptions, recite class features, and walk you through character creation step by step. What it can't do reliably is apply the rules across a long combat. Spell slots get tracked inconsistently. Conditions stack incorrectly. The order of operations during a turn shifts depending on what's narratively convenient. Claude knows the rules in the way a player who read the books once knows the rules. It cannot enforce them like a game engine can.

What's the best Claude prompt to play D&D and tabletop RPGs?

The prompts that work best give Claude a clear setting bible, full character sheets, a recent session recap, and an explicit instruction to ask before retconning anything. Tell it to roll dice openly and show the number. Tell it to track HP, spell slots, and gold at the top of every reply. None of this fixes the underlying problems, but it slows the drift. The honest answer is that no prompt makes Claude into a rules engine. Prompting is a patch, not a solution.

How much does it cost to play D&D with Claude?

If you use the Claude consumer app, you're paying a flat monthly subscription. If you use the API directly (which many Game Masters do for custom prompts and tooling), a typical four hour session with a built up context can run into the low single dollars. Long campaigns get more expensive over time. On Claude Sonnet 4.6, any request that crosses 200K input tokens is billed at long context rates of $6 per million input and $22.50 per million output, roughly double the standard rate. That's the cost cliff to watch.

Are Claude's dice rolls actually random?

Not really. Claude generates plausible looking numbers from its training data rather than calling a true random number generator. A peer reviewed 2026 academic study of 11 frontier LLMs found that almost none of them passed standard randomness tests, and there's no way to verify during play that rolls aren't being shaded toward whatever outcome the narrative wants.

Is there an AI Game Master that doesn't have these problems?

Yes. Purpose built platforms like DungeonsDeep.ai combine a language model for narration with a separate rules engine, a real RNG, persistent campaign memory, and a structured character sheet system. The AI handles the story. Code handles the game.

Come Find Us

If you're curious about a real AI Game Master built for D&D and other tabletop RPGs, you can sign up for our beta at dungeonsdeep.ai. Otherwise, come say hi:

This work includes material from the System Reference Document 5.2 (“SRD 5.2”) by Wizards of the Coast LLC, available at https://www.dndbeyond.com/srd. The SRD 5.2 is licensed under the Creative Commons Attribution 4.0 International License, available at https://creativecommons.org/licenses/by/4.0/legalcode.

Back to all posts