Lessons: a topic goes in, a full lesson comes out
One sentence in. A complete, curriculum-aligned, illustrated lesson out, in a single click.
A real deck from the pipeline: illustrated, on-brand, curriculum-aligned, with images only where they teach.
~14 slides + worksheet + answer key from one prompt. ~$0.30 all in for a fully illustrated lesson. $0 for a repeat topic’s artwork.
The idea that makes it work
The lesson format is fixed. The model never chooses the shape of the lesson, only fills each slot’s content. That one decision turns “generate a lesson” from a black box into a pipeline I can engineer stage by stage, and measure.
The pipeline
A single reasoning call plans the whole lesson: structure, key points, image intent, the worksheet. Then a dozen cheaper models write the individual slides in parallel, each with the full-deck context and a word budget. Images are generated per slide, but only where an image actually teaches: a science topic gets a labelled diagram, maths stays clean, and some slides get nothing. Finally one renderer assembles it, and the same renderer drives both the downloadable deck and the live preview, so what you see is always what you get.
Decisions made on purpose
I put the expensive model where the leverage is, the single plan, and the cheap model where the volume is, the dozen parallel writers. That is cheaper and better than one mid-tier model doing everything. Every image is generated once, ever, and stored in a tagged library keyed to a style-locked prompt, so the second lesson on a topic pays nothing for its art. And illustration blocks: the lesson returns finished or not at all, because a half-built lesson is not a lesson.
How I measured “done”
Finished means every planned image is generated, stored and placed, or generation is not done. Every image logs its outcome, so a failure is a line I can read. When a real production incident hit, an unprovisioned storage bucket silently dropping images, I diagnosed it from the logs in minutes, not guesswork.
What I would bring to your product
Orchestration, not just prompts. Cost engineered in from the start through model tiering and caching. Frontier models shipped safely with validation, observability and a human in control. And the whole thing built end to end, from problem to live product.
