Inspiration
I keep hundreds of generated and exported design drafts. Before every review I open them one by one to find the versions worth showing, and generic-looking or broken screens still slip through. I maintain Taster, an open-source rule set that names the visual tells of "AI slop" design. I wanted that review to run where the files already live, on the NAS, without sending unreleased work anywhere.
What it does
Taster Design Review reviews the interface screenshots, mockups and exported frames in one folder or one date range. For each image the NAS vision model answers 23 concrete yes or no questions (purple gradient, gradient-filled headline, three identical icon cards, clipped or overlapping text, placeholder values). Fixed thresholds turn the answers into one verdict: keep, fix or slop. After the user confirms, the Skill binds a tag to each file, saves a Markdown critique, and can create an album of the keepers. It never deletes, moves or renames a file.
How we built it
It is an orchestration-only Skill written to the Uliya Skill Development Standard with the UGOS Skill Creation Tool. One SKILL.md delegates six self-contained steps to three registered experts: file-agent (ug-filesystem find, ug-filemgr tag-add and tag-add-file), photo-agent (ug-photo search-picture, vision-ocr, vision-vlm, create-album) and knowledge-agent (ug-knowledge save-text). The checklist and the report layout are Markdown references linked from SKILL.md. The package passes the kit's static check with zero errors and zero warnings.
Challenges
A vision model looks but does not measure. I limited the checklist to questions that can be answered from a single still image, made "unsure" never count toward a verdict, and marked contrast and overflow answers as advisory. The vision action takes one image per call, so a run is capped at 20 images.
Accomplishments that we're proud of
Every verdict is explainable: it cites a rule ID and what was seen. The review leaves a durable result on the NAS (tags and a report) instead of a chat answer that disappears.
What we learned
Writing for registered experts forces precise hand-offs: each step may only use the real result of the step before it, and an empty result must end the run instead of widening the search.
What's next for Taster Design Review
If selected as a Finalist: tune the thresholds on real model output with a set of known before and after screens, verify tag binding and album creation on the iDX6011 Pro, and record the demo. we ran into
Built With
- markdown
- ocr
- ugos
- ugreen-ai-nas
- uliya-skill-kit
- vision-language-model