Test versionNo API key is set up, so the language model is switched off. Requests are read by a built-in keyword reader and levels are built by the planner. Every level on this page was still built and checked by the real program.
NeuroBridge.SI Baku · AI Gaming track

Type a sentence. Get a puzzle level that can be won.

A tool for puzzle-game level designers. It learns a game's real rules from its code, builds the level you ask for, and wins it in the real game before handing it over.

Watch it workCheck your own levelSource code

A hard level, won by the computer

The problem

Nobody knows if a new puzzle level can be won until someone sits down and plays it.

WhoLevel designers at small puzzle-game studios and solo developers, who ship 50 to 200 hand-made levels per game.
TodayEvery level is drawn by hand and then played by hand to check that it works. Shortcuts and dead ends that slip through reach the players.
With Level-ForgeAsk in plain words. Get a level in the game's own file format, with its solution, proof that it needs the skills you asked for, and a measured difficulty.

Watch it work

Pick a request on the left. Each one was run through the real program and recorded. Press Play solution to watch the computer win, or Play it yourself and use the arrow keys.

Check your own level

Change the level and press Verify. The solver runs live in this browser tab and returns the shortest solution, or shows that there is none.

Level

UPPERCASE = a word tile · lowercase = an object · . = empty. Words: IS YOU WIN STOP PUSH SINK DEFEAT. Things: BABA KEKE ROCK BOX WALL HEDGE FLAG STAR KEY WATER LAVA SKULL.

Result

Two games so far

Baba Is You is only our example. The same steps apply to other 2D puzzle games, and we tested a second one. 3D puzzle games would need extra work and are not tested yet.

A generated Baba Is You level drawn with the game's own graphics
Baba Is YouThe rules are words lying on the board. Push the words and the rules change.
BABA IS YOUFLAG IS WINWALL IS STOP
100%of 200 requests gave a winnable level
A generated Sokoban level drawn with the game's own graphics
SokobanPush every box onto a target. Added in about an hour, with no change to the main program. 96.7%of 60 requests gave a winnable level

Both pictures are levels our program made, drawn with each game's own graphics. Baba Is You was made by Arvi Teikari (Hempuli). We used baba-is-auto, a free open-source copy of its rules by utilForever, and sokoban-solver-generator by xbandrade.

Side by side

How it works

Part 1 happens once per game and produces a rule book. Part 2 runs for every request and uses that rule book at every step.

InputThe game's codeand the levels that ship with it
→
Part 1 · once per gameLearn the rulesAn AI coding assistant reads the code, writes down each rule with the line it came from, and tests each rule in the real game. Words the game ignores are marked.
→
OutputRule bookrules · file format · level checker · solver
▼ every step below uses the rule book ▼
1UnderstandWhich skills? How hard?
2Possible?If not, refuse and say why.
3BuildSeveral levels in the game's file format.
4PlaySearch finds the shortest solution. The real game replays it.
5Fix or labelRepair what failed. Say what was proven.
In: your sentencePart 2 · runs for every requestOut: level file, solution, proof

Where the AI is

Coding assistant (Claude)

Read the 4,000-line game engine and wrote the rule book: 30 rules, each with the line of code it came from and a test in the real game.

Search program

Tries every possible move to decide whether a level can be won, what the shortest solution is, whether a skill is really needed, and how many traps there are.

Language model (Google Gemini)

Written to read free-form requests in step 1. It is not running in this test version because no working API key is set up. A keyword reader does step 1 in every result on this page.

Test results

All numbers come from scripts in the repository. None were typed by hand.

200 requests · three ways of making a level · one checker

What went wrong, and what we did

    Compared with today

    By handAsk a chatbotLevel-Forge
    Opens in the gameyesnot checkedchecked by the game's loader
    Can be wonafter play-testingno evidencesolution replayed in the game
    Shortcutsfound by lucknosearched for
    Difficultya feelinga guessmeasured
    Impossible requestfound latenot checkedrefused with the reason
    Time per levelnot measured by ussecondsseconds

    We did not time human designers, so we make no claim about hours saved.

    The rule book: each rule tied to a line of code and a test in the real game

    Show all rules
    Words that exist in the game's files but do nothing (requests for them are refused)

    Run it yourself

    On Linux or macOS

    git clone REPO_URL
cd level-forge
./build_engine.sh      # once, about 2 minutes
./run_demo.sh          # http://localhost:8765

    Without an API key the program runs in this same test version and says so at the top of the page.

    What it costs

    • Data: none to collect. It only needs the game's own code.
    • Computer: one ordinary computer, no graphics card. Most levels take under a second. Hard ones take about 15 seconds.
    • Language model: optional.
    • New game: one rule book. The second game took about an hour.

    Next

    • Get the language model running and measure it.
    • Add a third game with no human help.
    • A plug-in for level editors.
    • A trial with a real puzzle studio.