Arcanum

LitRPG Roleplay With AI (2026): How to Make Stats, Levels & System Windows Actually Stick

· by Rukka

There is a specific kind of AI roleplay that almost works and then falls apart in a way that is genuinely maddening. You set up a LitRPG campaign — levels, a status window, a skill tree — and the first hour is fantastic. By session three you are level 40, you have 8,000 gold you never earned, a skill you were granted has quietly vanished, and the numbers in your status box have no relationship to anything that happened.

LitRPG roleplay means playing a campaign where the game mechanics are visible in the fiction — status windows, experience points, levels, skills, and quest logs, in the tradition of LitRPG novels, progression fantasy, and manhwa “system” stories. It is the most rule-heavy form of AI roleplay there is, which is exactly why it breaks in ways ordinary narrative roleplay never does.

Almost everything written about AI and LitRPG is aimed at people writing LitRPG novels. This is about playing one — and it’s grounded in the engineering we’ve done on our own stat-driven engines, where these problems had to be solved for real.

Why AI Is Both Perfect and Terrible at This

The tension is worth understanding, because every technique below follows from it.

What AI is superb at: the flavour. Naming a skill Cinderstep and describing what it feels like to use. Writing the notification text when a System announces itself. Improvising a dungeon, a rival, a quest hook. Generating the texture of progression fiction is the single thing language models do best.

What AI is genuinely bad at: being a database. A model does not store your character sheet. It predicts plausible text — and your stat line is just more text. So each turn it produces something that looks roughly right rather than recalculating from a stored value. That’s not a bug you can prompt away entirely; it’s what the technology is.

The whole craft of AI LitRPG is arranging things so the model has to do as little arithmetic-from-memory as possible.

The Four Failure Modes

Name them and you can design against them:

  1. Stat drift. Your Strength was 14. Three turns later it’s 16, with no level-up. Nothing granted it — the model simply regenerated a plausible number.
  2. Reward inflation. Every fight grants more experience than the last, because the model matches rewards to narrative stakes rather than a fixed economy. You outpace your own world in days.
  3. Phantom skills. A skill you were granted stops being mentioned and effectively ceases to exist — or the reverse, the model invents one you never earned and starts using it.
  4. Quest amnesia. Open objectives vanish. The quest log is regenerated from whatever’s in recent context, so anything older silently drops off.

All four share one root cause: state that lives only in the conversation gets re-improvised every turn. Everything below is a way of moving state somewhere more solid.

Step 1: Specify the System Before You Play

The single biggest predictor of whether an AI LitRPG holds up is whether you defined the system before the first scene. A model with no economy invents one, and an invented economy always inflates.

Write these down and put them in your system prompt:

SYSTEM SPEC
Attributes: STR, AGI, VIT, INT, WIS (start 10 each, hard cap 99)
Level: starts 1. XP to next level = current level x 100.
XP awards: trivial 5-15 | standard 20-50 | hard 60-120 | boss 150-300
  No single award may exceed 300 without an explicit story reason.
Attribute points: +3 per level, allocated by the player, never automatic.
Skills: granted only on level-up, quest completion, or an explicit
  in-world trigger. Never granted mid-narration.
Currency: copper/silver/gold, 100:1. A day's unskilled labour = 8 copper.
Death: HP 0 = incapacitated, not dead, unless the player opted into permadeath.

Those specific numbers matter less than their existence. “A day’s labour is 8 copper” is what stops the model handing you 500 gold for clearing rats — it now has a reference for what money means.

Step 2: The Status Block Does the Real Work

This is the technique that fixes the most for the least effort. Require the model to reprint a complete status block at the end of every single turn, without exception.

It looks like decoration. It isn’t. Reprinting forces the current state into the most recent context on every turn, so the model is reading its own last-known values rather than reaching back through a long history. Skip the block for five turns and those five turns of values were effectively invented.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 KAEL VARN — Level 4 Skirmisher
 XP 210/400   HP 38/45   MP 12/20
 STR 13  AGI 17  VIT 12  INT 10  WIS 11
 Unspent attribute points: 3
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 SKILLS
 • Cinderstep (2) — short blink, 4 MP
 • Read Weakness (1) — reveals one enemy resistance
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 QUESTS
 ▸ The Long Silence — reach the relay tower (0/1)
 ▸ Debt to Mera — repay 40 silver (12/40)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 PURSE  1g 12s 40c
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Two rules that make it hold: every number that changed must be justified in the narration that turn, and nothing changes silently. If XP went up, the text says why. That single constraint eliminates most drift, because the model can’t quietly adjust a value it also has to explain.

Step 3: Tags Beat Raw Numbers

This is the technique from our own engine work that transfers most directly, and it’s counterintuitive.

When we built Aevum Realm Architect — a deterministic kingdom-builder where the entire game is an economy — the breakthrough wasn’t better arithmetic instructions. It was abstracting numbers into discrete tags. Your domain is defined by Domain Tags (farms, watchtowers, a scribe’s office) and your forces by Retinue Tags, and Projects add or remove tags with concrete effects.

Why it works: a model tracking “you have a Watchtower and a Granary” is doing categorical reasoning, which it’s good at. A model tracking “security 61.5, food 208.3” is doing arithmetic bookkeeping, which it’s bad at. Tags don’t drift the way decimals do, because there’s no plausible-adjacent value to slide toward — you either have the Granary or you don’t.

Apply it to LitRPG: prefer Fire Affinity II over fire resistance 34%. Prefer Wounded (−2 AGI until rest) over a running injury total. Keep raw numbers for the few things the genre needs visible — level, XP, HP — and make everything else a tag.

Step 4: Keep an Authoritative Copy Outside the Chat

This is the player-side workaround for a gap on the product side: across 28 scored products, Mechanical Depth is the weakest of our seven axes, and why AI RPGs can’t do mechanics covers what the two exceptions do differently. The lesson from every stat-driven engine we’ve built: giving the model something to consult beats asking it to improvise from a single instruction block. 8-Bit Kingdoms starts every campaign with a coherent kingdom for exactly this reason — reference files sit behind the prompt. Note what that does and doesn’t buy: it gives the model a solid world to consult, not a memory of your play, which is why the same game still loses the thread over a long dynasty.

For your campaign that means:

  • Keep the authoritative sheet somewhere that isn’t the conversation. Our free Campaign Memory Tool is built for this — world state, characters, and open threads tracked outside the chat and re-pasted when needed.
  • Add a correction command. Aevum ships a /fix_state command that snaps the game back to the last known good state when it drifts. Define your own: /resync followed by pasting your authoritative status block, with an instruction that the pasted block overrides anything the model believes.
  • Prune aggressively. Aevum automatically retires stale NPCs to keep long campaigns stable. Dead threads and forgotten side characters consume context and produce contradictions; cut them.

Step 5: Plan for the Long Campaign

Progression fiction is about the long haul, which puts LitRPG on a collision course with the context window. Status blocks make this worse — that box is reprinted every turn, and it’s pure overhead.

Which Model, and Where to Play

Model: system-style play rewards instruction-following above all else — Claude first, ChatGPT close behind, with more freewheeling models drifting on exactly the rule-tracking this genre depends on. The full breakdown is in best LLM for roleplay; no need to repeat it here.

Ready-made stat-driven games, if you’d rather play an engineered system than build one:

  • Aevum Realm Architect — deterministic kingdom-builder with an always-on HUD, a tag economy, and outcomes earned by preparation rather than dice.
  • Star Freighter Drift — a space-trading RPG with a real economy, which is progression-by-ledger in a different skin.
  • 8-Bit Kingdoms — generational dynasty management, where consequences outlive the ruler who caused them.

Building your own: the free AI RPG Prompt Builder assembles a structured system prompt, and the best RPG prompts covers the general craft. For packaging a ruleset as a reusable engine rather than pasting it each time, Claude Skills for tabletop RPG is the durable approach.

The System Apocalypse Angle

One subgenre deserves a note, because it suits AI unusually well: the “system apocalypse” premise, where a System descends on the modern world and announces itself with notifications, quests, and status windows.

It works because it makes the mechanics diegetic. In a normal LitRPG the status block is an interface layered on top of the fiction. In a system apocalypse it is the fiction — the System is a character, its notifications are events, and its rules are the plot. The model has to produce mechanical output anyway; this premise turns that output into story instead of interruption.

It also gives you a clean narrative lever for rebalancing. If the numbers have drifted somewhere silly, the System can issue a recalibration notice, and it reads as intentional.

Frequently Asked Questions

Can AI run a LitRPG or system-style campaign?

Yes, and it is one of the more rewarding things to do with an AI game master, but it is also the hardest form of AI roleplay to keep consistent. The narrative side works out of the box. The numbers do not: a model will quietly drift your stats, inflate your rewards, and forget skills it granted you unless you give it an explicit system to follow and a status block it must reprint every turn.

Why does the AI keep changing my stats and levels?

Because a language model predicts plausible text rather than maintaining a database. Your stat line is just more text to it, so each turn it regenerates something that looks roughly right instead of recalculating from the last known value. The fix is to make the numbers structural: define the rules up front, have the model reprint a full status block every single turn, and keep an authoritative copy of your state outside the chat.

What is a system window in AI roleplay?

A system window is the boxed status readout from LitRPG and manhwa progression fiction, showing level, experience, attributes, skills, and quests. In AI roleplay it does real mechanical work rather than being decoration, because forcing the model to reprint the full block every turn is what stops the numbers drifting. If the block only appears occasionally, the values between appearances are effectively invented.

Which AI model is best for LitRPG roleplay?

System-style roleplay is the most rule-heavy form of roleplay there is, so it rewards the strongest instruction-followers: Claude first, ChatGPT close behind. The model has to update numbers consistently and respect its own mechanics every turn, which is exactly where more freewheeling models drift. Our best LLM for roleplay guide covers the full comparison.

How do you stop the AI inflating LitRPG rewards?

Set the economy before you play and write it into the rules. Define what a normal experience award looks like, cap how much any single encounter can grant, and state that levels require a fixed and increasing threshold. Without those numbers the model has no reference for what is generous, so it escalates to match the drama of the scene and you hit level 50 by the third session.

Can you play a system apocalypse style campaign with AI?

Yes, and it suits AI unusually well, because the premise gives the model an in-fiction reason to be a rules engine. When a System announces itself to the world, status windows, quests, and notifications become diegetic, so the mechanical output the model has to produce anyway is part of the story rather than an interruption of it.

LitRPG, progression fantasy, and system apocalypse are genre descriptions, not products. Individual novels and series in these genres are the property of their respective authors and publishers. This guide is about running your own original campaigns — it does not reproduce any published work’s setting, characters, or system.