A JSON schema is a better model-to-Figma interface than a plugin API

Elliot Silver has published Compact Design, a small JSON language that a language model can write and a Figma plugin can import as native frames, with validation, normalization, lint and patch handled by a core package called @compact-design/core ^[1]. The Figma plugin is only the first adapter, and the core carries no Figma dependency at all: ordinary TypeScript, no figma.* calls, no plugin typings ^[2][3].

The ordering is the point. Input goes schema, then normalize, then semantic checks, then lint or patch, and only then to an adapter ^[4]. Schema and semantic errors block import; lint does not ^[5]. So there are two gates that stop bad output before it reaches the canvas and one that merely complains afterwards ^[20].

The surface a model has to hit is small. A node requires w and h, type defaults to FRAME, and IDs are generated when you omit them, though anything you intend to update later needs an explicit one ^[8]. Within that, the language still covers hex fills, HUG/FILL/FIXED sizing, angle gradients, elevation presets, variables, styles, components, prototypes and images, against a single current schema rather than a set of versions ^[9].

What makes this usable in a generation loop is the shape of the failure. Issues come back structured with severity, code, a JSON path, a message and a suggestion; the documented example points at nodes[0].children[2].layout.direction and states that the value must be HORIZONTAL, VERTICAL or GRID ^[6]. That is not a pass/fail signal, it is a repair instruction with coordinates. Silver's framing is that you should not ask a model to speak Figma's plugin API, and should not ask it for CSS and then treat a screenshot as a design ^[17].

The advisory tier is honest about its own limits. Lint flags contrast, 44px touch targets, 4px spacing, missing Auto Layout, hard-coded colours and broken prototype destinations ^[7], which means a schema-valid but badly built screen still lands. There is also no browser thumbnail; Silver argues a thumbnail would lie about fonts, Auto Layout, masks and effects, so the plugin returns counts, matches and repair JSON instead ^[10].

Round-tripping is where most model-to-canvas tooling falls apart, and this one at least has a mechanism. Every imported canvas and layer stores its JSON id as plugin data ^[11]. Create new refuses a canvas ID already on the page, Update matching IDs replaces canvases in place, Replace matching canvases rebuilds them, and Export selection writes Compact Design JSON back out ^[12]. Full-document update is canvas-granular while a patch is node-granular, and failures roll back because the previous design is cloned first ^[13]. Patch operations are set, remove and append, with set touching only the fields you list ^[14].

Caveats are the author's own. The package is not on npm yet, so you clone the repo, npm install, npm run build, then import plugins/figma/manifest.json through Figma desktop's development menu ^[15][16]. Silver's closing claim, that this is the format he would trust for model-generated UI, is his alone ^[18].

Watch for a second adapter, which is the only real test of whether the language is tool-neutral or quietly Figma-shaped ^[2]. Watch whether any lint rules get promoted to blocking ^[5][7], and whether "one current schema" ^[9] survives the first breaking change once other people depend on it.