Platma.

The builder had evolved into an expert tool: powerful, dense and hostile to first-time users, most of whom left before publishing anything. The AI assistant suffered the blank-prompt problem — users didn't know what it could do, so they didn't use it. Apps nested four levels deep (app → page → component → logic) with no consistent navigation model, and visual building and prompting competed as two separate mental models instead of reinforcing each other.

I ran task-based tests with no-code beginners to find the exact moment people got lost — usually the first empty canvas. I restructured navigation around a single persistent tree, then designed the assistant as a contextual collaborator: suggestion chips tied to the current selection, previews of what a prompt will change, and inline diffs to review AI edits before applying them.

The empty canvas became a guided start: templates, example prompts and a visible assistant instead of a blank grid. The AI panel shows its capabilities through context-aware suggestions rather than documentation. One navigation tree now anchors the entire builder, and AI output lands as reviewable changes — never silent mutations — which made users trust it with bigger asks.