Designing a human–AI content production system
A draft-first workflow for a children’s-book museum: agents research and generate structured drafts, while editors review, revise, and manually publish content.
The museum itself: a visual-first archive of children’s-book illustration and craft.
Building a visual archive with AI as a production partner
The museum is a visual reference for illustrators and designers, organized around craft: medium, technique, style, and influence.
I used the project to explore a second question: how far could AI participate directly in product design and implementation? Designing in code with the product context, working UI, and codebase in one environment created a tighter loop between design reasoning and implementation.
Concepting, design, and implementation collapsed into one continuous loop.
AI could generate content, but not earn trust
Structured metadata, such as author, year, and ISBN, was easy for AI to generate and validate. Open-ended research on medium and technique was less reliable: outputs could be vague, inconsistent, or difficult to verify.
A closed knowledge base improved grounding, but introduced a manual copy-paste bridge back to the CMS. The design problem was not simply how to generate content, but where AI could be trusted and where human judgment needed to intervene.
Experiment A — an AI refill button inside the CMS form.
Experiment B — a closed knowledge base, grounded but disconnected from the CMS.
Agent produces, interface validates, human decides
Rather than treating AI as a feature inside a CMS, I designed it as an agent operating across the production workflow. A custom Claude Code Skill researches a book, synthesizes findings, and generates structured content as a draft.
The workflow moves from research to draft generation, human review, admin editing, and manual publishing. This preserves speed without allowing unreviewed content to become public.
The workflow makes source grounding, draft status, and human publishing control explicit at each handoff.
The agent handles research, synthesis, and draft production. The admin interface makes drafts visible in context for review and correction. Human judgment remains responsible for resolving ambiguity and publishing approved content.
The same pipeline supports specialized skills—such as illustrator research and cover color-palette extraction—without a separate orchestration layer.
The review experience is still evolving. The project revealed that agent output needs a purpose-built validation flow—not just a standard CMS form—to make editorial review fast and reliable.
My involvement
Solo, end to end. I defined the product concept, designed and built the experience directly in code, tested AI/CMS workflow models, and developed the Claude Code Skills behind the draft-production pipeline. My focus was designing where automation should act, where human judgment should intervene, and how the handoff between them should work.