# Why Creative AI Projects Must Reopen Cleanly and Survive Targeted Edits

Published: 2026-10-18 | Domain: https://agentstudios.si | North Star: Working Files Over One-Shot Blobs

## Outcome first: preserving unchanged components across creative revisions

A one-shot generative video or a monolithic 2,000-line AI script looks impressive until a creative director asks for a targeted edit: *"Keep the fox's gallop cycle and the bridge physics exactly as they are, but change the pond reflection to twilight and add a lantern on the pier."* In an unassisted workflow, asking the model to edit the file frequently breaks the skeletal animation or deletes a collision check. In `agentstudios.si`, projects are stored as a modular solution graph of pinned working files so unchanged components are never regenerated during a revision.

## Three rules for revision-safe creative projects

1. **Component-level hash locking**: Every mesh generator, skeletal rig (`fox.glb`), shader material, and physics controller is addressed by its content digest in the project manifest (`project_graph.json`). When a revision targets only the lighting slot (`studio.lighting.four_pass_pbr.v1`), the runtime keeps the existing geometry and physics vertices byte-identical.
2. **Non-regression verification across revisions**: After applying a targeted edit, the harness runs the full project conformance suite (not just a check on the edited module) to verify that frame times, watertight manifold counts, and collision boundaries did not regress.
3. **Clean reopen in a fresh environment**: Every exported project bundle includes its pinned ES module import map, asset digests, and deterministic seed so it reopens identically in a clean browser profile or local workstation without hidden session state.