Claude Sdk System Prompt

System prompts in the Claude Agent SDK

The SDK accepts the system_prompt field in three different shapes. Pick by use case, not by what looks shortest:

File mode — the right default for production

Inline string prompts grow until they're unmaintainable. Move to a file as soon as the prompt is more than 5 lines:

system_prompt={"type": "file", "path": "my_agent/prompt/system_prompt.md"}

Benefits:

• Edit the prompt without touching code
• git diff shows prompt changes clearly
• Reviewers can read it without scrolling through Python
• Markdown rendering in editors makes structure visible

The SDK reads the file every time the option is constructed, so you can edit it and restart the process. Hot-reloading mid-session requires rebuilding ClaudeAgentOptions and starting a fresh ClaudeSDKClient.

The claude_code preset

system_prompt={
    "type": "preset",
    "preset": "claude_code",
    "append": "You're a SQL specialist. Prefer CTEs over subqueries. Always validate schema before writing migrations.",
}

This gives Claude:

• The claude_code system prompt (proactive, tool-using, error-handling personality, autonomy norms)
• Your append text added at the end

Use this when your agent should act like Claude Code — multi-step planning, tool use without permission-asking, terse status updates. Don't use it when you want a different personality (a customer-service voice, a teaching voice, a strict review voice). The base personality is opinionated and bleeds through.

Dynamic prompts — rebuild per turn

The SDK doesn't support "dynamic system prompt at runtime" as a first-class feature (no InstructionProvider like ADK). The workaround is to rebuild ClaudeAgentOptions and start a fresh client per turn:

async def run_turn(prompt: str, user_id: str):
    options = ClaudeAgentOptions(
        system_prompt=build_dynamic_prompt(user_id),
        # ... other options ...
    )
    async with ClaudeSDKClient(options=options) as client:
        await client.query(prompt)
        async for msg in client.receive_response():
            yield msg

You lose session continuity (each turn is a fresh session) — that's the cost of dynamic prompts. For multi-turn dynamic agents, use the SDK's session_id + resume to thread the conversation across these per-turn rebuilds.

A common middle ground: keep the base system prompt static and inject per-turn context via a UserPromptSubmit hook (see claude-sdk-hooks). Faster, simpler, no session juggling.

setting_sources and how skills layer in

options = ClaudeAgentOptions(
    system_prompt=...,
    setting_sources=["user", "project", "local"],
)

This loads CLAUDE.md files and .claude/skills/ directories from those layers — exactly like Claude Code does. user means ~/.claude/, project means the project root's .claude/, local means the current directory.

The system prompt you set via system_prompt is composed with whatever Claude finds in those sources. So you don't have to embed your skills' content in the system prompt — drop SKILL.md files into .claude/skills/<name>/ and they're discovered.

For nuvel-scaffolded SDK agents, setting_sources=["project"] is the sensible default — it picks up the project's .claude/ without leaking the user's home-directory settings into a server deployment.

Writing a good system prompt

Three sections, in order:

1. Identity — who the agent is, what it's for, who's talking to it.
2. Operating principles — concrete behaviors. "Use tools to take action, don't ask permission." "Cite tool outputs." "Surface ambiguity in one sentence; don't block."
3. Guardrails — things to never do, in declarative form. "Never delete user data without explicit confirmation."

Avoid the three patterns that hurt prompts:

• Long lists of "do this" — Claude weights the last item. Pick 3-5 principles, ordered by importance.
• Telling Claude its capabilities ("you can use tools, you can read files") — the SDK already tells it. Redundant text dilutes the prompt.
• Personality without purpose ("you are a friendly assistant who loves to help") — describes vibes, not actions. Replace with verbs.

Common mistakes

• Inlining a 200-line prompt as a Python string. Use file mode.
• Using claude_code preset for non-Claude-Code agents. The personality is too strong; users feel like they're talking to a coding agent.
• Dynamic prompts via Python f-strings rebuilt every turn. Works but breaks session continuity. Use UserPromptSubmit hook injection for per-turn context, or session_id/resume to bridge.
• Hardcoding tool names in the prompt. Claude already sees the tool list; mentioning specific tools in the prompt creates drift when you add/remove them.

Quick reference

# Static, file-based (recommended)
system_prompt={"type": "file", "path": "prompt/system_prompt.md"}

# Claude Code personality + your additions
system_prompt={"type": "preset", "preset": "claude_code", "append": "..."}

# With user/project/local skills auto-discovered
ClaudeAgentOptions(
    system_prompt={"type": "file", "path": "..."},
    setting_sources=["project"],
)