Hi everyone,
I am currently developing a complex, client-side CAD canvas layout engine that utilizes local execution loops via Gemini Nano. To facilitate our development, we have been leveraging AI Studio’s workspace environments. However, the current implementation of System Instructions (SI) is introducing a massive structural flaw that has repeatedly destroyed our active code builds during multi-turn evolution phases.
The Problem: Monolithic, Detached System Instructions
Right now, AI Studio treats System Instructions as a universal, static text block tied directly to the chat/prompt panel. While this is sufficient for generic text summarization sandboxes, it is completely inadequate for complex, multi-layered engineering projects.
Because the instructions are completely severed from the actual file directory, the model suffers from severe context displacement several turns deep into a session. When asking for a code refactor or an optimization sweep, the model routinely drops its formatting rules and defaults back to historical pre-trained weights (such as attempting to inject deprecated, legacy alpha parameters that silently break our modern production initialization paths).
We have had to completely reset and rebuild our workspace states numerous times because the global instruction box falls out of sync with the evolving code files. For 99% of amateur or independent builders who lack low-level systems knowledge to debug these silent model drops, this creates an impassable development barrier.
The Solution: A Repo-Scoped Instruction Manifest (e.g., AGENTS.md)
AI Studio needs to transition from a global UI text block to a project-scoped reference manifest. Instead of forcing developers to manage a detached text window, AI Studio should feature a configuration setting that points natively to a markdown or JSON file within the project’s own directory (e.g., AGENTS.md or .ai-instructions).
-
A Single Source of Truth: The AI’s build instructions travel with the repository. If a developer modifies a token-pacing helper or updates an invariance rule in the files, the workspace agent reads the update automatically.
-
Pre-Processor Behavior (Ignored on Deployment): Similar to a
.gitignoredirective, this instruction manual would actively steer the AI workspace engine during construction and iteration phases, but would be automatically stripped out upon final deployment—keeping the production token payload lean, fast, and secure.
If AI Studio is going to become a robust ecosystem for building real, deterministic web applications, the instructions must be fundamentally anchored to the code itself. I would love to hear if the engineering or product teams are considering repository-scoped manifests for future workspace updates.
Cheers,
Joseph