Picture books tell their stories through pictures, but good writing about the illustrations was scattered and scarce in Chinese. Working solo as designer, developer and editor, I built a bilingual picture-book museum for Taiwanese readers, then an AI research workflow inside its admin. I review every paragraph before it’s published, and research time dropped from 7–11 hours to about 2–3 hours per book.
AI Studio gathers evidence and drafts inside the admin, so one curator can keep publishing.
33 of 71 books are published, each with sourced visual analysis.
A review queue keeps a human approval step on each paragraph.
I wanted to understand how the pictures tell the story. The material was scattered across interviews, exhibition pages and articles; Chinese-language writing on illustration was thin; and references I saved on Facebook and Instagram were hard to find again.
Each close reading needed research, visual study, sourcing, writing and fact-checking, about 7–11 hours per book. With 38 books waiting and a team of one, writing alone couldn’t keep up.
Each entry focuses on how the images tell the story, with visual analysis. A timeline gave historical context but broke down at around 20 books, when titles from the same year overlapped. I prototyped three homepages; Option A, one featured book a day, gave visitors a reason to come back, so the museum became a destination worth researching for, before any AI was added.
Per-field generation wrote generic commentary about friendship and childhood instead of reading the pictures. Splitting the work between an admin advisor and a Cursor skill improved the research, but I still copied every draft in by hand.
Analysis of illustration and page layout
General commentary, not visual analysis
I merged both into AI Studio inside the admin, where AI drafts with the whole book and its sources in view. The biggest gain came from removing the copy-paste between tools, not from a better prompt.
Every AI paragraph links to numbered evidence, a source or a visual observation. Claims with only one source are flagged, not cited, so I can check them before anything goes live.
I tested with 5 readers: parents, picture-book fans and people who promote picture books. They wanted more visual analysis and help talking with their kids, so I added close-read mode and an age guide. Requests for post-reading activities shape what’s next.
Research time dropped from 7–11 hours to about 2–3 hours per book, so one curator can keep the museum growing.
Next: I choose the angle of analysis before AI writes, my photos of the physical book feed the visual analysis, and research gaps get flagged instead of producing failed drafts. I’m also bringing editing into AI Studio, so I can revise and approve each paragraph right in the draft without going back to the book page.
Visit the museum