Inspiration
A folder full of assets is not yet a usable scene. A game developer might find a promising character model but still need to know which textures accompany it, whether the documentation describes the required rig, and what remains unverified. Filename search alone cannot answer those questions.
ForgeVault grew from my work on AI-assisted game-asset tools. I wanted a creator's NAS to support a more useful question: "What belongs in this scene—and what can the available evidence actually tell me?"
What it does
ForgeVault: Scene Asset Flightcheck is a Round 1 UGREEN Skill designed to turn a creative brief into an evidence-backed virtual scene kit.
The workflow starts with an authorized folder, required and excluded asset roles, technical constraints, and privacy preferences. It searches within explicit limits, inspects supported metadata and readable documentation, and produces:
- A role-by-role asset shortlist.
- Source-linked evidence for its recommendations.
- Documented missing companions and unresolved dependencies.
- Conflicting or unknown facts that need checking.
- Prioritized next verification steps.
Original assets remain in place. Optional report saving and project tagging require approval of the exact effects and targets.
The distinguishing idea is to evaluate asset relationships and evidence, not merely similarity. A high preference score cannot override permissions or hard constraints, and missing information does not silently become a confident recommendation.
How I built it
I authored the Skill and its core workflow for this hackathon using the official UGREEN developer package. The submitted ZIP contains SKILL.md and six directly linked Markdown references.
The Skill orchestrates Uliya's registered experts: file-agent for discovery, metadata and approved tagging; document-agent for documentary evidence; and knowledge-agent for approved report persistence. It uses the user's permitted model configuration for synthesis. A local-only request must stop if that processing mode cannot be verified.
Separately, I built a pure-PHP research lab containing 128 role/context-specific rankers, each stored in a 500-byte parameter file. Directly programmed weights, synthetic fitting, audits and reproducibility checks explored compact preference models. This experiment informed the design, but the rankers do not execute inside the submitted Skill and are not NAS dependencies.
OpenAI ChatGPT and parallel assistant workers assisted with research, design, authoring, testing and review. No previous Skill or model checkpoint was reused. The official UGREEN kit was consulted under its applicable terms and is not redistributed; the Skill ZIP includes no third-party library, model or asset.
Challenges and lessons
The hardest problem was deciding when the system should not make a claim or take an action.
A bounded search cannot prove an asset is absent. A sidecar file can contain misleading instructions. Documentary compatibility is not the same as successful engine import. An earlier approval may no longer match the current files.
ForgeVault therefore separates supported, inferred, unknown and conflicting evidence; discloses search coverage; treats source text as untrusted; and requires approval and readback for optional writes. It also distinguishes a completed advisory report from an unverified future operation.
Evidence and limitations
The package passed the official static checker with zero errors and zero warnings. In an initial 22-case offline advisory evaluation, all 31 checked candidate decisions matched, while 19 of 22 exact workflow-state labels matched. Those disagreements remain recorded. Six later targeted guard cases met the reviewed invariants.
These are package checks and offline assistant evaluations—not proof of NAS runtime reliability. The separate PHP experiment is synthetic and does not establish real-world asset quality.
ForgeVault does not claim binary-mesh inspection, engine import, rig validation or legal clearance.
What's next
On an assigned UGREEN NAS, I will validate the real expert interfaces, every cited source, privacy behavior, bounded reads and approved per-file tag outcomes. I will also test capped discovery, malicious sidecars and stale approvals.
The goal is a trustworthy production shortlist that helps creators move from "files I own" to "a scene I can confidently prepare."
Built With
- markdown
- ugos
- uliya
