Inspiration

I'm a songwriter in Lagos with over 100 song recordings on my phone, saved as voice memos with names like "New Recording 47". Good hooks get lost, I record the same idea twice, and finding "that line about rain" means replaying everything. I also don't want unreleased songs going to a cloud transcription service. A NAS that runs AI locally is the right home for them.

What it does

Hook Vault is a UGOS Skill with two modes.

Build mode

  1. Finds the voice memos in a folder you choose, reads their length and format, and spots duplicate takes.
  2. Transcribes only the memos that are new since the last run.
  3. Gives each idea a type (Hook, Verse, Melody, Freestyle or Note), a working title and up to three theme words. Humming with no words becomes a Melody idea instead of invented lyrics.
  4. Shows you a table to approve or edit, then, with your approval, tags the memos and saves a dated index next to them.

Find mode

  • "Find the idea where I sang about rain": answers from the index, with the matching lyric lines as evidence.
  • "Show me all my hooks": lists memos by tag.
  • "Play that one": plays it if the memo is in the NAS music library.

It never moves, renames or deletes your memos.

How I built it

With UGREEN's UGOS Skill Creation Tool (uliya-skill-kit 0.7.3). The Skill is an orchestration-only SKILL.md that hands each step to UGREEN's registered experts: file-agent (search, media info, duplicate scan, tags), knowledge-agent (transcription and saving text), document-agent (questions over the index) and media-agent (playback). Every step passes only real results from the step before, and every delegation state (completed, failed, cancelled, waiting for the user, waiting for approval) is handled. It passes the kit's static checker with 0 errors and 0 warnings. I used Claude (Anthropic) as a drafting assistant for the Skill text and the proposal.

Challenges I ran into

Designing for honesty: sung lyrics transcribe less accurately than speech, so weak results are flagged "check by ear" instead of hidden, and empty results are reported as they are. Making repeat runs safe was the other: only new memos are processed, and every run saves a new dated index instead of overwriting the old one.

What I learned

How to think in delegations: the Skill decides what should happen, and the NAS experts do the real work, with approvals built in.

What's next

Validate it on a real UGREEN AI NAS with my own recordings: transcription quality on sung vocals, the tagging flow and playback.

Built With

  • claude
  • markdown
  • ugos
  • ugreen-nas
  • uliya-skill-kit
Share this project:

Updates

Submission history