Ask an AI to write a lesson, then a quiz, then slides, then a script, and you get four artifacts that slowly disagree. Change one and the others don't know. Compilers solved this problem for software with an intermediate representation (IR).
The problem with independent generation
Each output is generated from the source separately, so small differences creep in. When the source changes, every output has to be regenerated and re-reviewed from scratch.
The Learning IR concept
Represent the course once, as typed objects: objectives, claims (with source links), lessons, assessment items, scenarios, media cues and locales, each with a stable ID. Every delivery format is compiled from that one model.
What it enables
- Consistency: the quiz and the lesson can't disagree, because they reference the same claim.
- Incremental updates: a changed claim updates only the objects that use it.
- Testability: checks run on structured objects, not prose.
- Portability: SCORM, xAPI, cmi5, QTI, web and documents from one release.
Where humans matter
The IR doesn't replace instructional design. It gives designers one place to make decisions that every format respects. Coursecast's Learning IR is on the roadmap; its first form is the JSON export you can use today.