Dungeon crawls and text adventures with an AI game master
A text adventure is a game first. The narrator has to say no, keep a map, and let you die, which is everything an assistant model is tuned against.
What a dungeon crawl needs
Dungeon crawls and classic text adventures are about exploration and resource management: rooms with exits, locked doors, traps, monsters, an inventory that matters. The model has to keep a consistent map, respond to commands literally, refuse actions that are impossible, and apply consequences, including death. Tone can be anything from deadpan Zork-style to grim, but the rules have to hold or nothing you do matters.
Why chat models make poor game masters
Assistant tuning makes models say yes. You try a locked door and it opens; you fall in a pit and land softly; the monster flees. The map rearranges itself because the model does not remember what it described. Filtered models also refuse or soften monster fights and character deaths, so the dungeon has no teeth. A text adventure with no possibility of failure is a guided tour.
The other problem is verbosity. Classic text adventures answer "open door" with one line. An assistant model answers with a paragraph about the creak of ancient hinges and what the adventurer feels. That gets slow over hundreds of commands, and long descriptions also bury the exits and items the player needs to see.
Making the dungeon hold together
- Write the map down. A simple room list with exits ("Room 3: guardroom. N to 4, W to 2, locked E door") in a lorebook entry. Generate it in a separate chat if you want it unseen.
- Keep inventory in the author's note. Update it yourself or ask the model to print it on request.
- Tell the model to say no. "Impossible actions fail with a brief reason" is the single most useful line.
- Use short replies. Room description, what changed, done. Text adventures are terse.
- Attach a map image. If you drew one, the vision-capable models can read it.
Sample game master prompt
You are the engine of a text adventure dungeon crawl. Respond to {{user}}'s
commands literally and briefly, in second person.
The MAP in the lorebook is fixed. Describe rooms and exits exactly as listed.
Impossible or failed actions fail, with a one-line reason.
Traps and monsters are dangerous. {{user}} can be injured and can die.
Track INVENTORY; items cannot be used unless held.
Keep replies under 120 words unless a fight is happening.Samplers and model choice
- Temperature 0.6 to 0.7, the lowest of any genre here. Text adventures need the engine to be predictable.
- Min-p 0.1.
- Repetition penalty 1.0. Room names and exits must repeat exactly.
outlaw-1 is a good engine for most crawls: cheap per turn, and it accepts a map image. If the dungeon has puzzles, such as levers, codes and doors that depend on each other, glm-5.3-outlaw reasons through them much more reliably.
Failure modes and fixes
- The map changes. The map entry is not in context. Make it always active.
- Everything works. Add the "impossible actions fail" rule and lower temperature.
- Purple room descriptions. Cap reply length and ask for a deadpan tone.
- Invented items. Inventory in the author's note, with "items not listed do not exist".
FAQ
What temperature should I use for a text adventure?
Low, around 0.6 to 0.7, with min-p 0.1 and no repetition penalty. The engine needs to describe the same rooms the same way every time.
Can the AI read a dungeon map I drew?
Yes, Outlaw 1, GLM 5.3 Flash Outlaw and MiMo V2.6 Flash Outlaw accept images, so you can attach a map.
Which model is best for puzzle-heavy dungeons?
GLM 5.3 Outlaw, the full model, is the strongest on the line at reasoning through puzzles and dependencies. Outlaw 1 is cheaper for simple crawls.
Related
