Inspiration

Every object around us is the answer to a design problem, but we only see the finished answer. A clothes hanger bends instead of snapping. A camera body is sealed against dust and light. Nobody tells you why. I wanted a way to point at an object in my own photo library and get the why: the forces, materials, physics and chemistry behind the design, in layers I can keep opening up.

What it does

Design Teardown Blueprint is a Skill for UGREEN AI NAS. From one real product photo on the NAS, it builds a layered, text-based design-study blueprint:

  • Level 1 (whole object): design intent, part map, how it works step by step, physics and chemistry notes, materials and manufacturing, trade-offs, and how entry, mid and high-tier components differ.
  • Level 2 and deeper (zoom): say "zoom into the shutter button" and get a deeper section for that part only: its job, working principle, why that material and shape, how it fails, and what to open next.
  • Optional: cite the user's own manuals and datasheets stored on the NAS, then save the blueprint as Markdown and tag the photo and blueprint, so studies become a searchable collection.

Every statement carries an evidence label: [Observed], [Read], [Cited], [Inferred] or [General]. The vision model only sees the outside of an object, so hidden parts are labelled as inferred and are never presented as seen.

How I built it

The Skill is pure Markdown, built with the UGREEN Skill Developer Package. A parent workflow delegates to registered NAS experts and never touches the NAS directly:

  • photo-agent: visual question answering and text recognition on the photo, and the zoom passes
  • document-agent: answers from the user's own reference documents, with citations
  • knowledge-agent: saves the blueprint (after platform approval)
  • file-agent: tags the photo and blueprint (after platform approval)

I used Claude (via Claude Code) to check the idea against the capability catalog, draft the Skill files, and run the kit's static checker until it reported zero errors and zero warnings.

Challenges

  • Staying inside real capabilities. The NAS cannot generate images, and the Skill has no web search. So the blueprint is structured text, and references come only from the user's own documents.
  • Honesty about what a photo shows. A photo cannot show a camera's sensor. The evidence labels exist so the Skill never claims more than it saw.
  • Safe workflow design. Each step uses only the real result of the previous one. The Skill stops on an empty search, an unidentifiable object, or a missing citation, and it handles all five delegation outcomes.

What I learned

Good AI explanation depends less on cleverness than on separating what was seen from what was inferred. Designing the evidence labels shaped the whole Skill.

What's next

This is an early build that has passed the static checker only. It has not yet been tested on a real UGREEN AI NAS. Next: real-photo testing (including an unidentifiable object and a part that isn't visible), then tuning the zoom depth.

Built With

Share this project:

Updates

Submission history