# Generate

An LLM writes the app as a set of files.

## Getting started

Use `generateObject` with a schema for the file tree, and pass the skills from
`@rivet-dev/dynamic-apps` as the system prompt so the model knows the supported
project layout:

```ts
import { anthropic } from "@ai-sdk/anthropic";
import { rivetActorsSkill, webServerSkill } from "@rivet-dev/dynamic-apps";
import { generateObject } from "ai";
import { z } from "zod";

const { object } = await generateObject({
  model: anthropic("claude-sonnet-5"),
  // Skills teach the model the supported project layout.
  system: [webServerSkill, rivetActorsSkill].join("\n\n"),
  schema: z.object({ files: z.record(z.string(), z.string()) }),
  prompt: "Build a team board.",
});

const { files } = object;
```

Pass the generated `files` to [`deployApp()`](/dynamic-apps/docs/deploy). If the
build fails, feed the diagnostics back to the model and deploy again. See
[Deploy](/dynamic-apps/docs/deploy) for the repair loop.

## Skills

Skills are prompt fragments exported by `@rivet-dev/dynamic-apps`. Each one
carries the instructions and a complete starter project for one kind of app, so
generated code compiles and serves on the first try more often.

| Skill | What the model learns |
| --- | --- |
| `webServerSkill` | The supported TypeScript Fetch-handler layout: `package.json`, `tsconfig.json`, and a Hono entrypoint with no application-owned listener. |
| `rivetActorsSkill` | The same layout plus Rivet Actors: durable state, actions, and HTTP routes backed by actors. |

More skills will be added over time.