System Prompt★

A system prompt is the baseline configuration block injected into a model request. Unlike user messages (which change turn by turn), the system prompt acts as the standing code of conduct. It defines who the model is, what styling conventions it must follow, which tools are available, and how it must handle errors or security decisions.

You configure system prompts when:

• Writing backend integrations to restrict Claude to outputting only valid JSON objects.
• Building code refactoring agents that must follow project rules.
• Enabling specific personas (e.g. Socratic reviewer, SQL debugger).

The Instruction Anchor

Technically, model providers process the system prompt with highest priority, aligning the model's attention heads to the system tokens throughout the session.

A standard system prompt organizes guidelines using structured XML or Markdown blocks:

1. Role Definition: "You are a senior TypeScript compiler assistant."
2. Constraints: "Never edit files outside src/components/. Never add external dependencies."
3. Tool Guidelines: "Always run npm run lint after modifying a file to verify formatting."
4. Format Constraints: "All code must be encapsulated inside markdown blocks."

Because the system prompt is processed on every turn, it spends a portion of your Attention Budget. Keeping it lean and using Progressive Disclosure is crucial to prevent performance slips.

Field Applications

1. Fullstack Developers (System Parameter Integration)

• Code Example:

  const completion = await client.messages.create({
    model: "claude-3-5-sonnet",
    system: `
      You are a database migration helper.
      CONSTRAINTS:
      - Output raw SQL queries inside \`\`\`sql blocks.
      - Do not output any conversational introduction or greetings.
    `,
    messages: [{ role: "user", content: "Add a timestamp column to the users table." }]
  });