Bench Log

The engineering notebook nobody writes: turns a week of bench photos, scope screenshots, voice memos and receipts on your NAS into a cited build log.

Comment

Inspiration

I build hardware, and I sit on the executive board of Build18, Carnegie Mellon's hardware hackathon. The real record of a build is captured with busy hands: a phone photo of a whiteboard, an oscilloscope screenshot, a twenty second voice memo, a crumpled receipt. Nobody writes the engineering notebook. By demo day the reasoning behind the design is gone, and at reimbursement time so are the receipts.

What it does

Bench Log is a UGREEN AI NAS Skill. You say something like "write up this week's log for my robot arm folder" and it:

  1. Inventories every capture in that folder for the date range (file-agent, ug-filesystem find)
  2. Transcribes voice memos (knowledge-agent, ug-knowledge audio-transcription)
  3. Reads bench photos and scope screenshots (photo-agent, ug-photo vision-ocr and vision-vlm)
  4. Extracts receipts and reconciles them with your budget (document-agent, ug-document extract and reconcile)
  5. Drafts a dated engineering log where every line cites the file it came from (ug-document writing)
  6. Shows you the draft to approve, edit, or cancel
  7. Saves the approved log and, if you want, creates a week album (ug-knowledge save-text, ug-photo create-album-condition)

It never fills gaps: an empty day reads "no record" and an unreadable value reads "needs review". Steps 1 to 5 are read only, and nothing leaves the NAS.

How I built it

The Skill is a pure orchestration SKILL.md built against capability catalog 2.0.0 from the UGOS Skill Creation Tool 0.7.3. Each step delegates to the single registered expert that owns the CLI, passes only verified upstream results, and handles completed, failed, cancelled, waiting_user, and waiting_approval explicitly. The package passes the official checker with zero errors and zero warnings.

Challenges

Designing honest failure behavior: partial results are kept and labeled, and a missing capture is reported, never invented.

What's next

If selected as a Finalist: test on a real UGREEN AI NAS with a synthetic Build18 project folder, tune vision prompts for scope screenshots, and measure transcription quality on noisy bench audio.

AI disclosure

Claude (Anthropic) helped research the capability catalog, draft SKILL.md and its references, and draft and edit the proposal and this description.

Built With

Share this project:

Updates

Submission history