Inspiration
AI agents are becoming increasingly capable of managing files through natural language. However, file operations can have unintended consequences, especially when multiple files are involved.
I wanted to explore a different approach: instead of focusing only on automating file management, what if an AI agent could make every proposed change transparent, require explicit approval, and verify the outcome?
This idea became ChangeGuard NAS, a skill designed to make AI-assisted file moves more accountable and easier to inspect.
What it does
ChangeGuard NAS provides a structured workflow for moving explicitly specified files on a NAS.
Before moving anything, it inspects the source files and destination folders, identifies filename conflicts, and prepares a detailed change plan.
Only files explicitly approved by the user may proceed. The skill then rechecks their states before requesting each move through the platform's registered file agent.
After each attempted move, it checks the source and destination and reports the outcome as verified, failed, not executed, pending approval, or unverified.
It does not automatically overwrite files, rename conflicting entries, or perform destructive cleanup.
How we built it
I developed ChangeGuard NAS using the official UGREEN Uliya Skill Starter Kit and Claude Code.
The project uses a Markdown-based SKILL.md to define the workflow and delegate operations to the registered file-agent, which accesses the ug-filesystem capability.
The design uses file metadata and SHA-256 hashes for inspections, the documented cpmv operation for moving files, and separate user confirmation and platform approval steps.
Three reference documents define the reporting format, technical limitations, and development-tool disclosures.
ChatGPT supported project ideation, requirements analysis, technical risk review, and proposal drafting.
Challenges we ran into
The biggest challenge was designing a reliable workflow within the documented platform capabilities.
The available documentation does not guarantee atomic moves, rollback, or the exact behavior of every file operation.
Another challenge was distinguishing a genuinely missing file from a file whose state could not be determined because of a permission error or failed inspection.
I addressed these limitations by narrowing the scope, blocking uncertain operations, rechecking file states before execution, and distinguishing verified outcomes from uncertain ones.
Accomplishments that we're proud of
- Created a complete, documented Uliya Skill for guarded file moves.
- Designed separate approval and post-operation verification stages.
- Added explicit conflict detection and failure-state handling.
- Passed the official static checker with zero errors and warnings.
- Packaged the Skill and its linked references into a validated ZIP structure.
The project has passed static validation but has not yet been tested on a live NAS.
What we learned
This project deepened my understanding of AI agent orchestration, file-system operations, authorization boundaries, and failure handling.
An important lesson was that a successful tool response does not necessarily prove that the intended operation occurred. Reliable automation requires examining the resulting system state.
I also learned the difference between passing static validation and demonstrating correct runtime behavior.
What's next for ChangeGuard NAS
The next step is to test ChangeGuard NAS on the UGREEN platform using controlled file-movement scenarios.
I plan to evaluate approval handling, filename conflicts, concurrent file changes, unexpected errors, and post-move verification.
Future improvements may explore more robust recovery mechanisms and broader file-management workflows, subject to platform capabilities and successful validation.