What AI GAME BUILDER can and can't do
AI GAME BUILDER is the AI inside the editor. It takes a prompt and directly changes the open scene and the project's files. This page explains how far the AI's changes reach, why a person has to check the result, and which tasks are better done by hand.
What the AI changes
AI GAME BUILDER doesn't stop at suggesting code or explanations. It changes the editor's actual data. In our tests, the AI did the following.
| Target | What the AI did (observed examples) |
|---|---|
| Entities in the scene | Creates, deletes and renames entities. In a new project's default scene, it renamed Plane to Ground and deleted Box (tested once). |
| Hierarchy | Groups entities under parents such as Environment, Lighting, Gameplay and UI. Group names vary from run to run. |
| Components | Adds render, collision, script and UI components to entities and sets their values. |
| Scripts | Creates new script files in ASSETS or edits existing ones. It added a Korean // description line at the top of each file (all 3 files we opened). |
| Materials | Creates materials named like Color E63946 in a Materials folder and shares them between entities of the same color. |
| Models | Downloads the models it needs from the asset store on its own, puts them in a Store folder and uses them. In all 3 runs tested on 2026-10-26, it made the stars from a star model in the store. It also succeeded all 3 times when asked to switch to a store model the user had downloaded. |
| Run checks | While working, it runs the game itself for about 1.5 to 4 seconds at a time to see whether errors occur and whether the controls respond. |
The AI works separately, then applies all of its changes to the scene at once when it finishes and leaves the line 씬에 반영했습니다 · 엔티티 N개 추가 … (Applied to scene · N entities added …). Applied changes are saved immediately, just like changes you make by hand. You don't need to press a save button. The AI asks you to confirm its plan once before it starts, but it doesn't ask again when the finished work is applied to the scene.
It asks for confirmation once, before starting
When you send a prompt, the AI first reads the current scene, writes up a plan, and then shows the 이대로 진행할까요? (Proceed as planned?) card. Click 예 to start the work. Click 아니오 (No) and the AI changes nothing and asks what you want changed. In tests on 2026-10-27 to 2026-10-29, this card appeared for every prompt we sent, from building the first game to fixes and model swaps.
This is the only confirmation. After you click 예, there is no further approval until the work is done, and then the changes are applied to the scene all at once. While the AI works, the send button turns into 중단 (Stop), and it changes back to the send button when the work is done.
The plan doesn't guarantee the result. In one case, the AI changed an entity during the work that the plan said it would "leave as is". What to read in the plan is covered in Read the plan and work steps.
Changes you didn't ask for
The AI sometimes adds or changes things that aren't in the prompt. These are usually changes meant to make the game more convincing or to fix problems the automatic check found. Here is what we observed when building with the same star-collecting prompt 7 times.
- It put up walls or railings around the edge of the ground (5 of 7 runs).
- It didn't delete the
*Probe.jsfiles it created for checks, and repurposed them for other features (4 of 7 runs). - It added a restart on the R key (1 run).
- It added decorative effects, such as floating stars or a blinking win message (1 run).
The same happened with prompts that only added a rule or asked for an explanation.
- When adding a lives rule, it deleted the 4 perimeter walls because the ball couldn't fall off otherwise (2 of 2 runs).
- The automatic check decided the scene was "too dark (brightness 9%)", so the AI changed every material in the scene from a night look to a day look and raised the camera (1 of 2 runs). The prompt ended with "Leave everything else as it is."
- The AI reported that it had lowered lives to 1 for testing and then restored them, but lives were actually still set to 1 (tested once).
- We asked only for an explanation of a script and added "Don't change anything," but the automatic check changed the color and size of the timer text (tested once). The script file we asked about stayed the same size.
Other scenes can change too
The scenes in a project share script files. So a task you give in one scene can change the game in another scene. In an experiment with 7 scenes in one project, another scene broke in 2 cases. In 1 more case, the AI couldn't create a new file because its name clashed with a file left by another scene. The structure is explained in What's per scene and what the project shares.
How to find the extent of the changes is covered in Check what the AI changed. The basic order is: read the change report, check the changed assets, and open the scripts.
Why a person has to verify
The AI also runs the game while it works. But each run lasts only a few seconds, and the number of runs per prompt is limited. So it's safest to treat the AI's check as covering only "it starts without errors and moves". In some cases it missed defects that only show up when you play the game to the end.
- In a run where the automatic check ended with
문제 없음(no issues), playing it ourselves turned up 3 defects. The win message was showing from the start, collecting a star didn't increase the count, and the camera didn't follow the ball. - In a run where a lives rule was added, the AI reported that it had "checked losing lives and game over". In fact, game over was showing right from the start.
- The AI once described a blue ball as a "red ball" (tested once). The code comment still had the old content.
Also, the AI only checks whether something "works". Whether 3 lives is too generous or whether the restart position is frustrating is something a person has to find out by playing. So after every AI task, run the game yourself with Launch. How to interpret the final line is covered in Read the last line and the auto-check.
What the AI can't do
These are the limits we confirmed in our tests.
- It can't play a game from start to finish. For the reasons above, a person has to find the defects that only show up at the end.
- It doesn't always finish a prompt in one go. Of 7 runs of the same prompt, 1 stopped partway, and 3 hit the run limit and couldn't finish the final check. What to do is covered in When the AI stalls, stops short, or loops.
- Ctrl+Z undoes only the AI's changes to the scene, not script files (tested on 2026-10-29).
Pending Recheck in the current editor whether image (texture) generation and sound checks are possible. In the previous version, it was observed that the AI couldn't draw images or hear sound
By hand or with the AI: which is better?
Editing by hand uses no credits, takes a few seconds, and changes only the field you edit. An AI prompt takes a few minutes, uses credits, and may change things you didn't ask for. On the other hand, tasks that require several entities and scripts to fit together are easy to get wrong by hand.
| Task | Better option | Why |
|---|---|---|
| Setting position, rotation or scale to exact values | By hand | Enter the values in the Position, Rotation and Scale fields in ENTITY. |
| Changing a color | By hand | Change the material's DIFFUSE → Color. Entities that use the same material may change along with it. |
| Changing a number in an existing rule (speed, number of lives) | By hand | Change an attribute value such as maxSpeed in the entity's SCRIPT fields. |
| When you don't know where that number is | Ask the AI, then change it by hand | When we asked where it was and added "Don't change anything," the AI named the exact entity and field in about 1 minute (tested once). Then make the change by hand. |
| Adding a new rule (lives, time limit) | AI | For one lives rule, the AI created 3 entities and 1 script, edited 1 existing script, and deleted 4 perimeter walls. It took 2 to 3 prompts and 10 to 30 minutes (2 runs). |
| Changing many entities at once | AI | You don't have to repeat the same change by hand for each entity. |
| Fixing a defect whose cause you don't know | AI | When we sent 3 defects as a numbered list, describing exactly what we saw, the AI found the causes and fixed them in 4 minutes. In another run, a similar prompt took 10 minutes 30 seconds. |