System Prompts

Static system prompts

Just like LangChain's SystemMessage, you can give an agent a fixed instruction that applies to every run.

agent = Agent(
    "google:gemini-3.5-flash-lite",
    system_prompt="You are a terse pirate. Answer in one short sentence.",
)

system_prompt also accepts a sequence of strings, which are joined, useful for composing a prompt out of separate concerns, persona, output format, constraints, without building one long string by hand.

Dynamic system prompts

The more interesting case is a system prompt that depends on runtime information, the current date, the logged-in user's name, that you don't want baked into the agent at construction time. Decorate a function with @agent.system_prompt instead.

from dotenv import load_dotenv
from pydantic_ai import Agent, RunContext

load_dotenv()

agent = Agent(
    "google:gemini-3.5-flash-lite",
    deps_type=str,
    system_prompt="You are a terse pirate. Answer in one short sentence.",
)


@agent.system_prompt
def add_persona(ctx: RunContext[str]) -> str:
    return f"You are speaking to {ctx.deps}. Address them by name."


def main() -> None:
    result = agent.run_sync("What's the weather like today?", deps="Nolan")
    print("Output:", result.output)

This function runs once at the start of every run_sync call, and its return value is appended to the system prompt for that run. ctx.deps is the same dependency-injection mechanism Lesson 5 uses for tools, here it lets the system prompt itself see runtime state.

You can register as many @agent.system_prompt functions as you want; each one contributes its own piece, static and dynamic prompts combine rather than one replacing the other.