Build a visual system, not one-off graphics
Before generating anything, decide three constants: a palette, an illustration style, and a layout convention for text space. Then derive every asset from them — course cover, module headers, lesson thumbnails, slide backgrounds — by reusing the same style phrases in every prompt. gpt-image-1.5 follows layout instructions like "clean left two-thirds for text, subject on the right" reliably, which matters when every slide needs room for teaching content.
Thumbnails that organize the curriculum
In a course library, lesson thumbnails function as navigation. A pattern that works: one consistent background style per module (so modules read as chapters at a glance), a distinct central icon or scene per lesson, and short lesson titles in bold type. Batch-generate a module's thumbnails in one session with a numbered template prompt, and students can visually locate "that lesson about pricing" without reading a single title.
Illustrate concepts, don't decorate slides
The highest-value educational imagery is explanatory: a visual metaphor for a difficult idea outperforms any decorative graphic. Generating "a dam holding back water, one small gate open, controlled flow" for a lesson on rate limiting teaches; a generic tech-blue abstract does not. Build a habit of asking, for each hard concept, "what image would make this obvious?" — then generate exactly that. Verify any labels or text in the image before publishing, and add real captions in your slide tool.