System prompts

Use systemPrompt for Runtime instructions. Put user, assistant, and tool conversation content in messages. For the separate generateText / streamText API, the option is named system.

Runtime usage

import { createRuntime } from "@yourgpt/llm-sdk";
import { createAnthropic } from "@yourgpt/llm-sdk/anthropic";

const runtime = createRuntime({
  provider: createAnthropic({ apiKey: process.env.ANTHROPIC_API_KEY }),
  model: "claude-sonnet-4-5",
  systemPrompt: "Classify the email. Do not answer it.",
});

const result = await runtime.chat({
  messages: [{ role: "user", content: "Subject: Invoice correction" }],
});

A nonempty request systemPrompt overrides the runtime default. An omitted or empty string falls back to that default. The same selection applies to chat, generate, stream, and processChat.

await runtime.chat({
  systemPrompt: "Classify billing emails as REVIEW. Return one word only.",
  messages: [{ role: "user", content: "Please correct my invoice." }],
});

Keep application-owned instructions in server code. Validate client-supplied messages before treating any of them as system instructions.

Anthropic system-role messages

Starting in LLM SDK 2.6.0-beta.2, Anthropic Runtime also preserves system-role messages from existing histories:

await runtime.chat({
  systemPrompt: "Classify the email.",
  messages: [
    { role: "system", content: "Do not answer the email." },
    { role: "user", content: "Please correct my invoice." },
  ],
});

The effective Anthropic request has:

{
  system: 'Classify the email.\nDo not answer the email.',
  messages: [{ role: 'user', content: 'Please correct my invoice.' }],
}

The effective explicit/default prompt is first, then distinct system-role instructions in their original order. Both are instructions; the first does not erase or override the second. Since 2.6.0-beta.3, text repeated between the explicit/default prompt and a system message is sent once. Intentionally repeated instructions appearing only in system messages retain their order. Prefer the explicit field for new application code.

The fix extracts instructions before removing system messages from Anthropic's conversation array. Typed messages, raw tool histories, streaming, non-streaming, and tool continuations use the same behavior. Caller messages are not mutated. Since 2.6.0-beta.3, null and unsupported non-text system content are ignored while valid text is retained, restoring tolerance for existing proxy inputs.

Compatibility with earlier versions

In 2.6.0-beta.1, Anthropic Runtime silently dropped role: 'system' messages. Explicit systemPrompt worked. Applications using only the explicit field retain their behavior; applications using system-role messages now apply those instructions. This can change an answer that previously ignored the application's rules.

The separate modern Anthropic provider already handled system-role messages. The Runtime fix does not rename its API:

import { generateText } from "@yourgpt/llm-sdk";
import { anthropic } from "@yourgpt/llm-sdk/anthropic";

await generateText({
  model: anthropic("claude-sonnet-4-5"),
  system: "Classify the email. Do not answer it.",
  messages: [{ role: "user", content: "Please correct my invoice." }],
});

Guardrail output validation

Delivering instructions is separate from validating a result. A prompt asking for JSON may still produce Markdown fences or malformed output. Use supported structured output, validate the returned schema, and handle provider or parsing failure explicitly. The SDK does not define a default respond decision for your application, and this feature is not the pending native Decisions API.