The Agent Skills Directory

[DYNAMIC_EXECUTION]: The SaveManager implementation in references/json-saves.md (both GDScript and C#) uses load(entry["scene_path"]) and GD.Load<PackedScene>(e["scene_path"]) to instantiate game objects based on paths stored in the JSON save file. If an attacker modifies the save file to point to a malicious .tres or .res resource, Godot will execute any scripts embedded within that resource upon loading.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process external data from the file system (save files in user://), which represents an attack surface for data poisoning.
Ingestion points: The load_game function in references/json-saves.md reads JSON files from the platform-specific user://saves/ directory.
Boundary markers: None are used in the provided code to isolate or validate the content of the save file before processing.
Capability inventory: The skill uses FileAccess for disk I/O and load() / instantiate() for dynamic object creation.
Sanitization: While the code performs JSON structure validation, it does not sanitize the scene_path values to ensure they are restricted to a safe whitelist of project resources (e.g., res://scenes/enemies/).
[COMMAND_EXECUTION]: The skill utilizes Godot's FileAccess and DirAccess classes to perform file system operations (create, read, write, delete). While these operations are targeted at the standard user:// data directory, they provide a footprint for file system manipulation based on user-supplied slot names.

godot-prompter-save-load