Vercel AI SDK skills
vercel/aiThe AI SDK repo's skills: one public ai-sdk skill for building with the SDK plus about a dozen maintainer-only skills for the monorepo.
Vercel Labs' Generative UI framework plus 32 agent skills that document each @json-render package: catalogs, renderers, devtools, codegen and more.
recommended
A maintained Apache-2.0 set of package references that helps an agent write correct json-render code. It is only worth installing if you use the library.
check first: installs packages · relies on an outside CLI → security details
json-render is a framework for AI-generated interfaces. You define a catalog of components and actions with Zod schemas, the model returns a JSON spec limited to that catalog, and a renderer displays it. Specs can stream progressively as JSONL. The repo ships renderers for React, Vue, Svelte, Solid, React Native, Next.js, Ink terminal UIs, Remotion video, React PDF, React Email and React Three Fiber, along with 36 prebuilt shadcn/ui components.
The skills are API references for an agent writing code against those packages. The core skill covers defineSchema, defineCatalog, SpecStream and the experimental composeSpec decision-model API. Codegen documents traverseSpec, collectUsedComponents, serializeProps and the CodeGenerator type for building exporters. Other skills cover devtools (with short React, Solid, Svelte and Vue adapter notes), directives, image, ink, jotai, json-render-shadcn, mcp and next. The repo also includes general skills: skill-creator, which has an init_skill.py scaffolding script, and remotion-best-practices, which keeps topic rules files on subjects like maps, GIFs and 3D. An internal technical-writer skill under .agents/skills has three reference docs. Most skills are a single SKILL.md of a few KB with no references, scripts or evals.
32 skills averaging about 5.7 KB of API-focused instructions. Nearly all are a single SKILL.md, with a scaffolding script in skill-creator, rules files in remotion-best-practices, and no evals.
Counted from the repository file tree. How we score depth.
This repo is a set of documentation skills for the json-render libraries, including Remotion best practices and a skill-creator guide. The skills mainly tell the agent how to write code and to install @json-render/* npm packages in your project. The skill-creator bundles Python scripts that create and package skill folders; those scripts, the packages' source, and most repo files were not available for this review.
npm install for @json-render/* packages and peer libraries in the user's project; other skills (core, mcp, react-pdf, redux, jotai, next) give similar instructions.skills/devtools/SKILL.mdThe repository (1685 files) on Nov 1, 2026. Not a full review: 1577 of 1685 files were not downloaded (large repository; we reviewed the skill folders); we read 0 of 3 code files; it relies on code outside the repository: skills/skill-creator/scripts/init_skill.py (listed in CODE FILES, full text not provided); skills/skill-creator/scripts/package_skill.py (listed in CODE FILES, full text not provided); skills/skill-creator/scripts/quick_validate.py (listed in CODE FILES, full text not provided).
NVIDIA SkillSpector static scan: 34 pattern hits (MCP Rug Pull 20, Agent Snooping 4, Rogue Agent 3, Privilege Escalation 2, Excessive Agency 2, Anti-Refusal 1). The credential and 'Data Exfiltration' hits are documentation examples: a Mapbox token placeholder, an AI Gateway key in sample server code, and fetch calls to example.com and lottiefiles in Remotion snippets. The 'Anti-Refusal' hit is an API contract ('custom evaluators must answer every offered question'), and the self-modification hits are skill-creator steps telling you to edit SKILL.md. The U+2060 hit is a single stray word-joiner inside a Mapbox style URL string in a code sample, not hidden instructions. 'npx' MCP rug-pull hits are ordinary CLI commands in docs.
A review, not a certification. Skills change after we check them — read the files before installing.
Apache-2.0, commercial use allowed.
Not really. These are reference skills consulted while writing library code. They have no loops, stop conditions or verification steps for unattended runs.
The AI SDK repo's skills: one public ai-sdk skill for building with the SDK plus about a dozen maintainer-only skills for the monorepo.
The official skills inside the shadcn/ui repo: a shadcn skill for adding, fixing and styling components, plus a Radix-to-Base-UI migration skill.
The Remotion team's twelve skills for building videos as React code: project setup, markup rules, maps, captions, studio, render and upgrades.
Vercel's official 9-skill pack: React and Next.js performance rules, composition patterns, UI and writing review, React Native, deploy and cost audit.
json-render is a Generative UI framework from Vercel Labs. You define a catalog of components and actions, the model generates a JSON spec that can only use those components, and a renderer displays it. It supports React, Vue, Svelte, Solid, React Native, Next.js, Ink, Remotion, PDF, email and 3D scenes.
Install the core package plus a renderer from npm, for example npm install @json-render/core @json-render/react for React or npm install @json-render/shadcn for prebuilt shadcn/ui components. The agent skills are in the repo's skills/ directory, one folder per package.
Most do not. The experimental composeSpec decision-model API runs a server-side evaluator through Vercel AI Gateway with AI_GATEWAY_API_KEY, and the Remotion maps rule needs a free Mapbox access token. The other skills only document npm packages.
We found no hidden network calls, credential reads or telemetry. The scanner hits are documentation text. The only executable code is skill-creator's init_skill.py, which scaffolds a new skill folder. One code example in the Remotion maps rule contains a stray invisible character.
shadcn/ui is a component library you copy into your project. json-render is a framework that lets a model compose UI from a catalog you define. Its @json-render/shadcn package ships 36 shadcn/ui components ready to use as that catalog.