How to Hand a Large Unity Project to GPT-6 Astra Without Letting It Change Unrelated Systems
Do not start with “understand my whole game.” Give Astra a project map, one objective, an allowed-change boundary, and an explicit inspection phase. Expand context only along dependencies that the task actually touches.
Do not start with “understand my whole game.” Give Astra a project map, one objective, an allowed-change boundary, and an explicit inspection phase. Expand context only along dependencies that the task actually touches.
Use a map before raw files
For a mature Unity project, the first artifact should be a project map: major systems, scene ownership, important ScriptableObjects or services, testing entry points, and the directories that are out of scope. This is more useful than a giant undifferentiated context dump because it tells the model what each part means before it sees implementation detail.
Separate read scope from write scope
The model may need to read more than it is allowed to edit. State that distinction explicitly. A combat change may require reading player stats, animation state, input, and networking, while only one combat script is writable. This reduces the chance that an agent “simplifies” a dependency it merely needed to understand.
Require an edit plan before tool calls
- Files and objects to inspect first.
- Files and objects it proposes to modify.
- Why each modification is necessary for the stated objective.
- Tests or playtests that prove the change works.
- Rollback point if the validation fails.
Reject convenience refactors during feature work
Large repositories tempt agents to normalize naming, move helpers, or consolidate code while touching a feature. Those edits increase review surface without proving the requested behavior. Defer architecture work to a separate task with its own tests and acceptance criteria.
Working template
OBJECTIVE: [one behavior] PROJECT MAP: [systems + ownership] READ SCOPE: [what you may inspect] WRITE SCOPE: [what you may edit] PROTECTED: [folders / scenes / systems] FIRST RESPONSE: inspection plan only; no edits. AFTER APPROVAL: smallest implementation + engine validation + changed-file summary.
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.