How to Keep GPT-6 Astra From Rebuilding Working Game Systems During a Bug Fix
Make the existing behavior part of the contract. Name the system as protected, provide the smallest failing path, forbid architecture changes, and require the model to explain why the proposed patch cannot be localized before it touches a broader layer.

Make the existing behavior part of the contract. Name the system as protected, provide the smallest failing path, forbid architecture changes, and require the model to explain why the proposed patch cannot be localized before it touches a broader layer.
“Do not refactor” is weaker than a protected-behavior contract
List the behaviors that are known good and must remain identical. For example: jump height, input bindings, save format, network event names, and animation transitions. A model can technically avoid a refactor while still changing one of those contracts; the behavior list makes regression visible.
Use a blast-radius ladder
Allow the patch to expand one layer at a time: first the failing function, then its owning component, then an interface, and only then architecture. To move up a rung, require evidence that the lower rung cannot explain the failure. This keeps “cleaner design” from becoming a reason to rewrite a working system.
Require a no-change explanation for protected files
After the fix, ask for a list of protected files that were inspected but not edited and why. This simple report exposes accidental edits and makes code review faster.
Rollback if the patch needs unexplained collateral edits
If one bug suddenly requires ten unrelated files, stop and return to the last verified state. Either the diagnosis is wrong, the system has hidden coupling worth mapping first, or the requested fix is actually an architecture task and should be planned separately.
Working template
BUG: [...] KNOWN-GOOD BEHAVIORS: [...] PROTECTED FILES / INTERFACES: [...] MAX INITIAL BLAST RADIUS: [one file/component] RULE: expand scope only with evidence that the smaller scope cannot explain the failure. VERIFY: original bug + protected behaviors.
Source and verification note
This guide was checked against OpenAI / Playco case study and updated Sep 5, 2026. Product behavior can change; current primary documentation should take precedence over this guide.