Atlas of Bretonnia
FREN

Workshop journal · · 7 min read

A whole map in four days: what a frontier AI changes for modding

Tried a free AI and came away disappointed? I get it. What I used to bring the Season of Revelation into Warhammer III is a different tool: an agent, which reads your files, runs the modding tools, checks its own work and fixes it. Here is what it did, what it can do for your mod, what it costs, and how to start.

The campaign map in game: the forests of Athel Loren, their roads and, beyond, mountains
In-game capture: the Season map in Warhammer III.

“I tried it, it was mediocre”

While showing this mod around, I heard four reservations, all fair:

  • you need to understand what the AI produces;
  • nobody wants the creative part turned into prompts;
  • free models give mediocre results;
  • and the good ones cost money.

The third one deserves a closer look, because it's true, and it hides the key point. A free AI in a chat window is someone who answers and then waits. What brought a whole Warhammer I map into Warhammer III in four days, from 20 to 23 September, is something else: a frontier model, used as an agent.

A chat answers, an agent works

In a chat window, you copy an error, the AI suggests a fix, you copy it back, you run it again, you come back. You are the AI's hands. An agent (Claude Code from Anthropic, Codex from OpenAI) makes that trip itself: it reads the mod folder, runs the tools, reads what they answer, fixes, and goes again until the proof is there. You judge.

Who makes the trip?A chat windowThe AIYouThe toolYou paste, copy back, rerun, copy the error back: everything goesthrough your hands.An agent: Claude Code, CodexRead the folderRun the toolRead the outputFixuntil theproof is thereIt shows you the proof:you judge, in game
In a chat window, you carry everything back and forth. An agent goes round the loop itself: it reads, runs, reads the output, fixes, and goes again until the proof is there. You keep the judgement.

And the gap between models is real. On Terminal-Bench 4.0, a test where the AI has to complete real tasks in a terminal, Claude Opus 5.5 succeeds at 66.4% (Anthropic); Gemini 3.8 Flash, a fast and cheap model, at 19.1% (Google's table, via Vellum). More than three times as many tasks carried through: that's the difference between “it let me down” and “it worked”.

What it did on this project

I'm not asking you to take my word for it. Here is what was done, and you can see it:

  • A whole map ported: 400 × 440 hexes, 61 regions (57 with a town), 26 provinces, read layer by layer from Warhammer I's files and realigned for Warhammer III.
  • Ten playable lords instead of Warhammer I's two, with their quests, their victories and, for the Red Duke, mechanics of his own: the lost dukedom, Abhorash's favour, the blood tax and blood oath, Richemont's hunt.
The ten playable lords, on the parchment of the showcase video
A frame from the showcase video: animation, layout and the site's fonts, made by the agent from the game's portraits.
  • 1,499 texts in French and English, with official names checked, and 50 event pictures.
  • Games played by the AI, thirty turns in a row without a script error (with Orion, Drycha and Kemmler), to flush out crashes before a player hits them.
  • The 43 voiced lines of the mini-campaign, found again in Warhammer III's sound banks.
  • Landmark icons drawn with code, in Creative Assembly's style.
  • This site, its ornate map and its guides, and even the mod's showcase video: editing, sped-up loading screens, opening and closing animations, three encodes for Discord.

What it can do for your mod

  • Tables: read, create and fill hundreds of rows without a single key typo, through RPFM's MCP server (MCP is the standard plug through which an agent connects to a tool).
  • Lua scripts: quests, dilemmas, mechanics, written, reread and tested.
  • Texts: translations, official names checked, clean .loc files.
  • Crashes: reading a report, a log, a dump, and finding the missing data.
  • The tedious stuff: converting, realigning, comparing, counting, tidying. Everything that steals your evenings.
  • Documentation: it keeps the log of what was done and why, for you as for the next session.
The Red Duke's script opened in RPFM
The script that builds the blood tax dilemma, written by the agent and opened in RPFM. The RPFM guide shows, table by table, how the Duke was made.

The community is already getting ready. RPFM has shipped an MCP server since version 5 and, since 5.0.6, documentation to help agents decode tables, written by ChaosRobie (release notes). Tools like Total War Mods Translator translate thousands of mod lines with Claude or GPT.

You stay the modder

An agent doesn't decide what your mod should be. You give it a goal, a criterion, sources and don'ts; it works and brings you proof; you judge, in game, against your reference. Then you go again: iteration is what makes the quality, and you are the one steering it.

The brief, then the loopBriefGoalthe campaign holds 30 turns in “AI playseverything” mode.Criterion0 crashes, 0 errors in the script_log.Sourcesthe project guide, yesterday's testreport.Don’tsdon't touch the startpos; warn beforelaunching the game.Proofthe exact command, the turn report, alog extract.Controlthe same test on yesterday's version,which held 30 turns.The briefThe agentworksThe proofYou judge,in gamego again: iteration makes the quality
A brief in six lines: what you want, how you will know it worked, based on what, what is forbidden, what comes back, and the control case that lets you conclude. Then the loop: you judge each proof, in game, and go again.

What you delegate is your call. For me, the judgement, the lore and every decision stayed human; the rest was delegated, and I had more time for what I enjoy: watching the map come alive, and deciding what story it tells.

What it costs

In September 2026, Claude Code comes with Claude Pro, at $20 a month before tax; Max starts at $100 for long sessions (pricing). Codex comes with ChatGPT Plus, at $20 a month; Pro starts at $100 (pricing). Roughly the price of a DLC a month, for a tool that works while you judge.

Start with a ten-minute task

On a dated copy of your mod, with the agent open in its folder, give it a real small task: “Change this unit's cost, show me the row before and after, and tell me how to check it in game.” Then judge it in game: if it's right, you've just won back your first evening.

The full sequence, from installing the agent to that first task, is in Working with an AI on a big mod, along with the method that held up on this project.

Everything else, step by step, with examples to copy, is in the “AI for modding” guide. And if you want to build the rest of Bretonnia with us, suggest your ideas: it's a community project, and this is when everything gets decided.

Sources

Suggest an ideaSee the map