When the AI gets stuck: a debugging checklist
Your agent keeps "fixing" the same bug and nothing changes. Here's how to break the loop, in the order that works.
Every vibecoder hits this. The agent says “Fixed!”, you playtest, and the bug is still there. Three rounds later the code is worse. Stop, and work down this list.
1. Is the code even running?
Half of “it doesn’t work” bugs are code that never runs.
Add a print at the very top of the script and at the start of the
function that should run. Playtest and tell me which prints appear.Check the container. Server Scripts run in ServerScriptService and Workspace. LocalScripts run in StarterPlayerScripts, StarterGui and the character.
2. Is it the version you think it is?
With Rojo, make sure rojo serve is running and Studio says connected. With two Studio windows open, the agent may be editing the other place.
3. Get the real error
Ask the agent to read the output log during a playtest. Red errors include the script and line. An agent guessing without the error is just rolling dice.
4. Shrink the problem
Make a new minimal script that does only [the broken thing], with
nothing else. Does it work on its own?If the small version works, the bug is in how the pieces connect, not the feature.
5. Reset the agent
Long conversations fill up with wrong guesses the agent keeps believing. Start a fresh session, describe the bug from scratch, and include the error and what you already ruled out.
6. Ask for the cause, ban the patch
Do not add pcall, task.wait or nil checks to hide this. Find why it
happens and fix it there.