Skip to main content

The structured_prompt Step

structured_prompt prompts the LLM for a JSON-formatted answer that matches a schema. Bosun renders your prompt with the task context, requests a structured response, validates the result, and makes it available to downstream steps.

Key characteristics of a structured_prompt step:

  • Schema-driven output: The returns field tells Bosun what shape the reply should have. Bosun converts that into a JSON Schema before calling the LLM.
  • Template support: Like every string field, the prompt accepts Tera syntax so you can inject inputs, previous outputs, or loop variables.
  • Typed outputs: When a primitive return type is used, Bosun unwraps the value wrapper for you, so outputs.<id> yields the plain string/number/boolean/array.

Choosing a return type​

You can keep things simple with a primitive return type or provide a full JSON Schema when you need strongly shaped data.

Primitives​

Set returns to one of string, integer, number, boolean, string_array, integer_array, or boolean_array. If you omit returns, it defaults to string.

Example:

  - id: shortlist_features
    structured_prompt:
      prompt: |
        List three feature highlights for {{ inputs.product_name }}.
      returns: string_array

The resulting array is available directly as outputs.shortlist_features for later steps.

Custom schemas​

When you need richer structure, replace returns with an inline JSON Schema snippet. Bosun forwards the schema to the LLM and preserves the entire JSON object in outputs.

  - id: release_notes
    structured_prompt:
      prompt: |
        Draft release notes for version {{ inputs.version }} with clear headings and bullet points.
      returns:
        type: object
        required: [summary, highlights]
        additionalProperties: false
        properties:
          summary:
            type: string
          highlights:
            type: array
            items:
              type: string

With this configuration, the LLM must return an object containing summary and highlights, making the result easy to feed into other steps or tooling.

Referencing the output​

Just like other steps, the response is stored under the step index and optional id. For example, {{ outputs.release_notes.highlights }} gives you quick access to the generated list inside subsequent templates.

Error handling​

Structured prompts stop the task when templating fails, the LLM errors, or the response fails schema validation—unless you set continue_on_error: true. With the flag enabled, Bosun records the failure (including any schema mismatch details) and keeps the task running so follow-up steps can inspect errors.<step>. Read Error handling for more context.