ExperimentPreview
Experiment: documents written by AI
What happens when language models write interactive documents — and why a strict, validated format is a good target for them.
AIFile formats
Language models are good at writing small web applications and poor at guaranteeing they are safe. A format with a strict validator turns that weakness into a workflow: the model writes, the validator decides, and nothing invalid reaches a user.
Two pieces
- An authoring prompt. A self-contained description of the IDOP 1.0 rules — files, manifest, Runtime API, what the validator refuses — that any model can follow. It is published at /developers/ai-prompt/.
- A connector for assistants. An MCP server through which an assistant can create a document in the user’s IDOP Cloud. The document is packed and validated by the same core as the reader before it is saved, and the assistant acts only with the user’s permissions. It is built and in preview.
Early observations
Models follow the structural rules well when they are stated plainly. The most common failures are exactly the ones the executable boundary forbids — inline event handlers, a script loaded from a CDN, fetch instead of the Runtime API — which is the argument for a validator that refuses rather than repairs: the model gets a precise error code and tries again.