Inspiration
I publish serialized illustrated stories as vertical reels. Every episode is 20 to 40 frames made in separate sessions, and small things keep drifting: a hoodie gets darker, a backpack appears that was never there, a hallway floor changes pattern. Each frame looks fine on its own, so I only noticed after rendering. In one 26-frame episode, 8 frames had continuity defects, which meant regenerating them and rendering the whole episode again. A film set has a script supervisor for exactly this. Independent creators do not.
What it does
continuity-supervisor reads my series bible and optional shot list from the NAS, turns them into a cited continuity ledger, and checks each frame of an episode with the NAS vision model. It reports four kinds of drift: breaks of a written rule; frames that disagree with how a character, prop or place looks in most frames (or in the previous episode); stray readable text when the bible forbids it; and frames with the wrong size. Every suspected drift is asked again in different words before it counts as confirmed. The result is a review sheet saved next to the frames, ending with a "majority baseline" that the next re-check and the next episode reuse. On request, confirmed frames are tagged continuity-drift and moved to a _reshoot folder after approval.
How I built it
With the official UGOS Skill Creation Tool (uliya-skill-kit 0.7.7). The Skill is orchestration only: it delegates to file-agent, photo-agent, document-agent and knowledge-agent using catalog capabilities (ug-filesystem find/cat/mkdir, ug-photo meta, vision-vlm and vision-ocr, ug-document qa, ug-knowledge save-text, ug-filemgr tags, file-workflow batch-transfer). Each step uses only the previous step's real result, and every write waits for platform approval. It passes the kit's static check with 0 errors and 0 warnings.
Challenges
Vision models give false alarms on dark or stylised frames, and a tool that cries wolf gets ignored. So findings need either a written rule or a clear episode majority, and then a second, differently worded question; frames where the two answers disagree go to a "needs a human look" list instead of being tagged.
What's next
Validate on a real UGREEN AI NAS with my own 26-frame episode as the answer key, then tune the checklist wording for the NAS vision model.
Built With
- claude-code
- markdown
- ugos
- ugreen-ai-nas
- uliya-skill-kit
