When the AI stalls, stops short, or loops
When an AI task stops partway, finishes with work left undone, or repeats the same check over and over, don't start over. Continue in the same conversation. This page covers the signs that tell you whether to wait, and the prompts we actually sent in each situation.
Before you start
- You need to know how to use the AI panel and the conversation list. Read What AI GAME BUILDER can and can't do first.
- The meaning of the last line is covered in Read the last line and the auto-check. This page covers what to send based on that reading.
Here is how often these situations came up when we built the game 7 times in new scenes with the same star-collecting prompt.
| Situation | Observed |
|---|---|
| Stopped partway through the task | 1 of 7 runs (stopped at about 6 min 30 s) |
Last line was 문제 N건 남음 (N issues remaining), including the run that stopped | 3 of 7 runs |
| Hit the game-run limit and couldn't do the final check | 3 of 7 runs |
| Repeated the same check (looping) | 2 of 7 runs |
Prompts that add one rule didn't finish in one go in either of two runs; each needed two or three prompts.
Do it in the editor
Only the AI can continue an AI task. What you do yourself is judge the current state, and check the result with Launch after it finishes.
Decide whether to wait
While the AI is working, the send button at the bottom of the AI panel changes to 중단 (Stop). Use the table below to judge the current state.
| What you see | State | What to do |
|---|---|---|
Lines like ✓ Ball 생성 (✓ Ball created) keep appearing, or the tool count in the progress line (AI가 작업하는 중 · 단계 2/4 … · 도구 N회 · m분 s초, "AI is working · Step 2/4 … · N tool calls · m min s s") goes up | Working. It's normal for HIERARCHY and the viewport not to change until it finishes | Wait |
A plan with an 이대로 진행할까요? (Proceed as planned?) card is shown, and the button has returned to send | Not stalled. It's waiting for 예 (Yes) or 아니오 (No) | Read the plan and click 예. Nothing happens until you click it (observed once). This card also appears for continue prompts |
The progress line has stayed at AI가 생각하는 중 (AI is thinking) for several minutes | It may be waiting for another AI task to finish (observed once) | Check that no AI task is running in another tab. Send AI tasks one at a time |
AI가 게임 실행 확인 중… (AI is checking the game run…) appears at the bottom left of the screen | The AI is running the game to test it | Wait |
자동 점검: 문제 1건 — 고치는 중 (Auto-check: 1 issue — fixing) | The AI is fixing an issue it found itself | Wait |
The button has returned to send, but there's no final summary and no 씬에 반영했습니다 (Applied to scene) line | Stalled | When it stalls |
There's no final summary, but there is a 씬에 반영했습니다 · … line | Finished without a summary. Two fix prompts ended this way, and the results were correct | Check with Launch |
도구 호출이 중단되어 미완료 (Incomplete: tool call interrupted) | Stalled | When it stalls |
The last line is 문제 N건 남음 | May have stopped short | Check with Launch, then see When it stops short |
| The summary says "couldn't finish", "remaining work", or "the actual fix hasn't been applied yet" | Stopped short (even if the last line is 문제 없음, no issues) | When it stops short |
A sentence like run_game 호출 한도(9회)에 걸려 … (hit the run_game call limit (9) …) | The AI is reporting that it couldn't do the final check. Not stalled | Check it yourself with Launch |
| Sentences like "hit the tool call limit …" or "continue …" | Stopped short. It used up the number of tool calls allowed for one prompt (2026-10-27, 4 times across 2 runs) | Check with Launch, then see When it stops short |
| The same check keeps repeating, files it already fixed get fixed again, and the game-run window keeps appearing | Looping | When it loops |
The time a task has taken is also a useful signal. These are the ranges we observed.
| Task | Time taken |
|---|---|
| Building the first game (runs that finished in one go) | 6 to 10 min |
| 3 fixes grouped by number | 4 min to 10 min 30 s |
| Adding one rule | 10 to 30 min across two or three prompts |
| One continue prompt | 6 to 10 min. About 2 min and about 7 min in the 2026-10-26 runs; about 6 min and about 11 min in the 2026-10-27 runs |
| Building the first game (2026-10-26 and 10-27 runs) | About 8 to 12 min |
If the button shows 중단 and lines are still being added, keep waiting even past these times. Click 중단 only to stop a loop.
Check after it finishes
- Click Launch. If a game tab is already open, refresh it.
- Click the game screen once, then play from start to finish.
- Write down what's unfinished exactly as you saw it. If it's still there after you continue the task, use it in Send fix prompts.
Let the AI do it
All of the prompts below were sent in the same conversation. Don't start a new conversation.
When it stalls
It stopped partway. Check the current state of the scene and keep building the rest.We sent this to the 1 run out of 7 that stalled. The AI found what was missing (the win message and a broken link) and built it, taking about 10 more minutes to finish. On 2026-10-26 we also sent it to 1 run that stopped without a summary at the start of step 2. It finished about 7 minutes later, but a defect remained: the score didn't go up and the game didn't end. That run had stalled while running at the same time as an AI task in another project. In that run, the first attempt failed with a network error because the browser connection dropped; sending the same sentence again got it going.
"Check the current state of the scene" is there so the AI continues from what it has already built in the current scene. We haven't tested what happens if you tell it to start over from scratch.
When it stops short
It's not finished yet. Check the current state of the scene and finish the remaining work. Leave everything else as it is.We sent this in the five cases below, and all five times the AI finished the remaining work.
| Case | Before continuing | Time taken | Result |
|---|---|---|---|
| First build (7th run) | 문제 1건 남음 (1 issue remaining). Win message shown right at start, count stuck at 0 | About 6 min | Passed |
| Adding a lives rule (first run) | 문제 1건 남음. Summary said "the actual fix hasn't been applied yet" | About 10 min | Lives went 3 → 0 and game over worked |
| Adding a lives rule (second run) | Summary said "couldn't finish all the way"; last line was 문제 없음 | About 7 min | Lives went 3 → 0 and game over worked |
| Different genre (dodging game) | Summary said "remaining work (important)" | Not recorded | The missing link was connected, but health didn't go down on collision |
| First build (2026-10-26, run that ended at step 3 due to the tool call limit) | 문제 1건 남음; summary listed "remaining tasks" | About 2 min | Finished, but the win message was shown from the start and the count was stuck at 0 |
As with the last line, "finished the rest" is not the same as "has no defects". Always run the game yourself after continuing. Send any remaining defects as a separate prompt in the format from Send fix prompts.
When the AI's summary left a check undone, saying something like "I'll verify it if you ask once more", we sent this:
Please do the remaining check. Run the game and confirm that it ends when all 10 stars are collected. Leave everything else as it is.In the 1 run (the 6th) where we sent this, the AI confirmed that the game ends after collecting 10 stars. After that, though, it kept repeating the same check. That case continues under When it loops below.
In summary, of the 7 first-build runs, the 3 we continued (the 3rd, which stalled; the 6th, which had a check left; and the 7th, which stopped short) all finished with the prompts above.
When it loops
In 2 of 7 runs, the AI didn't finish and kept repeating the same check. In one of them it repeated the same check 3 times over about 5 minutes. While looping, it also fixed files it had already fixed.
- Click 중단 at the bottom of the AI panel. The task stops and the button returns to send.
- Send the following in the same conversation.
Stop checking. If any temporary test code is still in place, revert it. Then summarize only the causes you found and what you fixed so far.When we tried this once, the task stopped as soon as we clicked 중단, and about 2 min 30 s after sending this prompt we got a summary of the causes found and the fixes made. That run passed when we ran it ourselves. This was observed once, so don't expect the same result every time.
- When you get the summary, run the game yourself with Launch.
- Use the steps in Check what the AI changed to check whether test files or temporary values remain after the summary.
- Send only what still doesn't work, in the format from Send fix prompts.
New conversation or continue
| Situation | What we did in testing |
|---|---|
| Handling a stall, stop-short, or loop | Continued in the same conversation (all cases above) |
| Fixing the game we just built, adding rules | Continued in the same conversation |
| Building a new game in a new scene | Clicked 새 대화 (New conversation) and started there (all 7 runs) |
Conversations in the AI panel stay the same when you switch scenes. When you start a new game in a different scene, click 새 대화 first.
Once, we asked for a rule to be added in an existing scene using 새 대화. The AI recognized the livesManager.js already in the project and kept using it. However, the first prompt ended in under a minute with "couldn't finish all the way", and we completed it by continuing with the "It's not finished yet…" prompt above. With only one observation, we don't draw any conclusion about how the new conversation relates to this result.
To go back to an earlier conversation, open the conversation list at the top of the AI panel. Each conversation's title is the beginning of its first prompt. Pick one from the list to continue it. If you click 새 대화 by mistake, pick the original conversation here too.
Rebuild in a new scene
If continuing doesn't help, you can create a new scene in the same project and start over.
- Click the scene name in the top bar to open the SCENES panel.
- Click + New Scene. An
Enter Scene name and press Enterfield opens below the list. - Once you see the field open, type a name and press Enter. If you press Enter before the field opens, the scene that was selected may open instead.
- In the AI panel, click 새 대화 and send your prompt again.
Building the first game again took 6 to 10 minutes. 3 of 5 rebuilt runs passed with no defects from the start. A new scene shares the project's scripts, so if the AI edits a shared script in the new scene, the game in the previous scene may change. How to check this is in Check what the AI changed.
Check the result
- When a continue or summary prompt finishes, reread the last line and the summary just above it.
- Play the game from start to finish with Launch. Some continued runs still had defects.
- If you stopped a loop, use Check what the AI changed to check for leftover test files and temporary values.
Troubleshooting
| Symptom | Cause | Solution |
|---|---|---|
| The button returned to send, but there's no final summary | The task stopped partway | Send the "It stopped partway…" prompt in the same conversation |
The last line is 문제 없음, but the summary says "couldn't finish" | Nothing was changed, so there was nothing for the check to flag | Send the "It's not finished yet…" prompt |
| It says it "hit the run limit and couldn't verify" | It used up its game-run limit (6 to 9 observed). It hasn't stalled | Check it yourself with Launch and send only what's wrong as a fix prompt |
A continue prompt failed with network error | The browser connection dropped (observed once) | Check your connection, then send the same sentence again |
| One feature still doesn't work after continuing | The remaining work finished, but a defect remains | Describe what you saw and send it as in Send fix prompts |
| The AI keeps repeating the same check | It's looping | Click 중단 and send the summary prompt |
| You clicked 새 대화 by mistake | The earlier conversation was only closed, not deleted | Pick the original conversation from the conversation list |