Nano Banana Pro Template Design Method

A workflow for designing Nano Banana Pro image-prompt templates using Cursor plus Claude Code (CC), which generates a first-draft prompt from global guidelines and runs the API directly in Cursor to view results.

An iterative visual-calibration loop: review the generated result, give targeted feedback on specific issues (e.g. 'the glass plane angle is wrong, it should always stay parallel to the canvas', 'fragments too small, narrative incomplete'), CC instantly updates the prompt, confirm the improvement and continue testing.

Compare a hard-light and a soft-light version to decide direction.

Working inside Cursor with the Claude Code agent.

Methodology writeup; core idea is writing general visual-design principles into /CLAUDE.md as global guidance (e.g. 'let visuals serve the narrative', 'principles over enumeration') so common methodology need not be repeated each iteration.

From concept to template (e.g. the 'memory fragment template' in figure 1): give CC a rough idea or reference, run several comparison tests until performance is stable, then a final trimming stage removes defensive redundant descriptions, keeping only necessary complexity and design preferences; the archived template is concise and transferable -- fill in the [Subject] variable to reuse it and the AI designs by context. Tip: make separate hard-light and soft-light versions and compare before committing to a direction.