THESK From Voices to Decisions

Inspiration

Meetings generate valuable information, but their recordings often remain unused after the discussion ends. Teams still need to listen back, write minutes, identify decisions, and turn commitments into actionable tasks. This creates friction between having a record of a meeting and actually using that record to move work forward.

THESK was designed to address that gap: turn recorded meeting audio into structured, reviewable meeting documentation.

What it does

THESK is an AI Meeting Intelligence Skill designed for UGREEN AI NAS. It orchestrates a workflow that:

Discovers audio recordings in a directory selected by the user.
Delegates audio transcription to the registered Uliya knowledge agent.
Organizes the available information into structured Minutes of Meeting (MoM).
Separates discussion points, confirmed decisions, tentative proposals, and action items.
Marks missing owners or deadlines as "[NOT SPECIFIED]" rather than inventing them.
Presents the complete draft for explicit human review and approval before requesting storage.

The goal is not simply to produce a transcript. It is to make meeting outcomes easier to review, understand, and act on.

How we built it

THESK was developed as an orchestration Skill using the UGREEN Uliya Skill Development architecture and documented capabilities.

Its workflow delegates file discovery to "file-agent" and transcription and storage operations to "knowledge-agent". The parent Skill coordinates these stages and enforces a human approval boundary before persistence.

The initial submission package contains a "SKILL.md" entry point and uses documented Uliya capabilities. It was checked with the official Skill validator as a static package validation step.

Challenges and design decisions

A major challenge was distinguishing what was actually agreed upon from what was merely discussed. A suggestion, unresolved debate, or silence must not automatically become a confirmed decision.

THESK therefore distinguishes confirmed decisions from tentative proposals and discussion points. It also avoids inventing action-item owners or deadlines, and marks unclear audio for human review.

Another challenge was designing around documented platform capabilities without assuming undocumented behavior. For example, the workflow requires an explicitly approved storage path and does not assume a default transcription speaker count.

What we learned

Building for an AI NAS requires more than connecting AI operations. A useful Skill must respect the platform's documented interfaces, handle uncertain outputs, make approval boundaries explicit, and communicate its limitations honestly.

THESK is currently a submission-stage Skill package. Static validation has passed, but physical UGREEN AI NAS runtime behavior, transcription quality, and performance still require device testing.

What's next

The next validation stage will focus on Indonesian and mixed-language recordings, noisy multi-speaker audio, empty directories, unclear speaker attribution, and storage or approval failures.

Our objective is straightforward: make the transition from meeting conversations to reliable, human-reviewed decisions and action items more structured and accountable.