1 of 15

Lab 5 – LLMize

Procedural Content Generation for Computer Games

Vojtěch Černý

cerny@gamedev.cuni.cz

2 of 15

Prerequisites

Previous HWs

3 of 15

The Task

Enhance one of your creations so far with LLMs!

It has to affect gameplay (!)

Attempt solving one of the big issues ☺

4 of 15

Codebase

Start with any of your previous HWs...

The most designed for are murder_mystery and roguelike

Both of the repos got an „llm“ branch

You can rebase your solution on top of it
(or just copy the last commit there)

5 of 15

Google Gemini

We‘re going to be utilizing Gemini models?

Why?

Probably the best free option right now
Watch out that your data may be used for LLM training

For Gemini-3.1-flash-lite (pretty small model, but useful)

15 requests per minute / 500 requests per day

(enough for some tasks, for some might need to split work to multiple days)

Gemma 4 is also decent, two variants, each get another 1500 requests/day

Gemini-3.5-flash (decent model all around, came out 2 days ago)

5 requests per minute / 20 requests per day

(can do final test in)

6 of 15

Google Gemini - technical

Need an API key (for free):

https://aistudio.google.com – get API key

For Python:

.env

GEMINI_API_KEY=xxxx

For Kotlin:

local.settings

GOOGLE_API_KEY=xxxx

7 of 15

Watch out for repetitions

LLMs want to repeat themselves

Why? LLMs are predicting text.

What text is likely to follow? The one that has already followed!

Beware of this when you want the llm to feel original or have choice...

This sneaks up when you don‘t expect it.

e.g. NPC chooses to be joking/serious for a question based on circumstance

But because he was joking on the question before (and it‘s in the context)

He will most likely be joking again... And again...

8 of 15

Watch out for not enough variability

LLMs want to repeat themselves

Why? LLMs are predicting text.

What text is likely to follow? The one that has already followed!

Beware of this when you want the llm to feel original or have choice...

This sneaks up when you don‘t expect it.

e.g. NPC chooses to be joking/serious for a question based on circumstance

But because he was joking on the question before (and it‘s in the context)

He will most likely be joking again... And again...

9 of 15

Options (for inspiration)

For roguelike:

Talking NPCs
LLM generating a high-level concept/story of the levels
Combat by describing player‘s actions
LLM generated effects in game (combining building blocks?)
...

For murder_mystery:

Talking to NPCs directly
LLM coming up with details of the puzzle
LLM adding new rules for each game
...

10 of 15

Options (for inspiration)

For mario:

LLM thinking like a designer and adding „intent“ into the level
...

For minecraft:

LLM designing objects that get spawned in biomes
...

11 of 15

Big issues to solve

LLM letting the player do anything they want
Characters saying out-of-place things
Consistency across time
Drifting away from design
LLM design and content are not in-sync
Multiple people sounding different each time
Characters being bad at lying
...?

12 of 15

Criteria

LLM enhancement of one of your HW – 1pt
Has effect on gameplay – 1pt

(meaning it‘s not just a „new“ UI for the game, like NPCs just chatting)

Solving one of the big issues – 1pt

13 of 15

Submission details

Submit via this form
Submit:

link to your fork of the project (or zipped)

Don’t forget to have comments in your code!

3 screenshots
describe your approach
describe what you have tried, what worked or didn’t, the problems you faced
how much time you spent with the project (will not be part of evaluation)

Deadline is 10.6.2026

If you need more time, talk/write to me before the deadline

14 of 15

Questions / Problems

You can always e-mail me at cerny@gamedev.cuni.cz
I’m also frequently online at Discord https://discord.gg/4nfAzG

15 of 15

Q & A

Vojtěch Černý

cerny@gamedev.cuni.cz