System Prompts

examples/19-system-prompts.halo

How to give agents persistent identity with system prompts WHY: System prompts define who the agent is — its persona, constraints, and output format — separate from what it does in each pipeline step.

Open in the playground →

halo

// ============================================================
// STEP 19: System Prompts
// ============================================================
// LEARN: How to give agents persistent identity with system prompts
// WHY:   System prompts define *who* the agent is — its persona,
//        constraints, and output format — separate from *what*
//        it does in each pipeline step.
// ============================================================

// Without system prompt — agent has no inherent identity
agent Generic [openai/gpt-4o-mini, temp: 0.5]

// With single-line system prompt
agent Translator [openai/gpt-4o-mini, temp: 0.3]
  system "You are a professional translator. Always respond with the translated text only, no explanations."

// With multi-line system prompt (triple-quoted)
agent Analyst [anthropic/claude-sonnet-4, temp: 0.1]
  system """
    You are a senior data analyst.
    Always respond with structured JSON containing:
    - result: the analysis output
    - confidence: a score from 0.0 to 1.0
    - methodology: one-line description of approach used
    Never include raw data in responses.
  """
  equips StripeAPI

tool StripeAPI : REST("api.stripe.com/v1", apiKey: Env.STRIPE_KEY)

// The system prompt is sent as role:"system" to the LLM.
// The pipeline step's prompt becomes role:"user".
// This mirrors how LLM APIs actually work:
//   messages: [
//     { role: "system", content: <agent's system directive> },
//     { role: "user",   content: <pipeline step prompt> }
//   ]

pipeline TranslateGreeting(name: String, lang: String) -> Result:
  name
    // The Translator's system prompt ("You are a professional translator...")
    // is automatically prepended. This prompt is the user message.
    |> Translator.prompt("Translate 'Hello {{name}}' into {{lang}}")
    |> System.Print()

// NOTES:
// - system belongs on the agent, not the pipeline step
// - If you need a different persona, declare a different agent
// - system prompts and policies are orthogonal:
//     system = shapes intent ("respond in JSON")
//     policy = enforces limits (MaxWords, Redact)

System Prompts ​

System Prompts