AI self-describing configuration for chemical visualization

Getting complex visualizations looking consistent across platforms, formats, and versions has historically been onerous and cumbersome. Figuring out the schema, thinking about how to make it work as versions change, making editors so that humans can actually make it look good for humans: this is a lot of work.

The ease that AI agents can both create good enough editors, and understand configuration between systems allows a new kind of artifact: a self-describing and editable configuration editor:

At the Open Molecular Software Foundation while working on unifiying visualization tools a sidequest had me modifying rdkit 2D renderings. I needed to edit the rendering of the 2D molecules, then apply those settings to a different code package.

Using framejs I used AI to quickly create a visual editor, and had the editor export JSON that I could then simply copy into the actual package I was working on. The editor described its own configuration, the AI agent can ingest the editor and configuration, so there is no need for an external canonical complex schema to unify, the editor provides its own.

This is not an argument against schemas, in fact, schemas are actually more useful with AI agents, as the surface area is easily understood by agents and schemas reduce the context between different agents. This is more like a schema-lite: a way to create consistent communication between agents, tools, and humans with visual needs, without over-designing or cumbersome separation and publishing of schemas.

For highly visual tools in molecular dynamics, quickly creating your own editor to configure visualizations that are meant to be shared across platforms, systems, and codebases is something good for us humans.

Example 2D chemical visualization settings editor, with JSON export and description:

See live example you can take and modify yourself