Instructions

What this page is for

The Instructions workspace section shows the exact instructions stored for a game. The main runtime instructions remain the canonical authoring surface for turn generation, additional instructions can be frozen into universe generation contracts for MGPT, and the generated scene image instructions separately guide Generated Scene illustration behavior.

Why MDL exposes the full instructions

• Instruction transparency matters. Game developers should be able to inspect the exact text guiding the runtime and image generation.
• Iteration is easier when the stored source and the runtime source are visibly the same thing.
• Validation can now talk about mismatches against real workspace content instead of a hidden composition layer.

What belongs in runtime instructions

• World and tone rules.
• Character behavior and relationships.
• How scenes, character visuals, music, and effects should be interpreted.
• Any stable gameplay constraints the runtime should follow every turn.

What belongs in generated image instructions

• Stable art direction for Generated Scenes.
• Rules for how character and location references should be interpreted.
• Composition guidance such as camera angle, continuity, lighting, and mood.
• Visual quality guardrails that should stay consistent across image generations.

Generated Scene recomposition

Generated Scene prompts should treat reference images as continuity inputs for a new illustration, not as finished plates. The goal is to keep the same location identity while staging a different shot: a changed camera position, crop, depth, foreground, or angle that fits the current story beat.

This matters because simply placing characters over an existing background reads as cheap and visibly AI-generated. The strongest Generated Scenes feel magical because they can stay in the same room, street, courtyard, path, or landscape while changing the scene entirely: reverse angles, closer framings, doorway views, foreground occlusion, side views, and partial landmark crops all preserve place identity without reusing the original view.

• Use enough visual cues for the place to remain recognizable.
• Let background details shift, recede, crop out, or be covered by characters.
• Prefer concrete camera and staging instructions over broad layout-preservation language.
• Judge success by whether the result feels like a newly illustrated beat in the same location.

What does not change here

MDL still owns the engine-level state shape and rendering semantics. Instructions can explain how to use those fields well, but the engine remains responsible for what fields like background, foreground, specialScene, specialBackground, specialCharacter, specialChoices, music, and sfx mean at runtime. The retired top-level keys scene, characters, description, and generatedScene should not appear in new instruction content.

Next