🔄

Multi-Step Workflows

Orchestrate complex multi-turn conversations with clear state management.

The Challenge

Multi-step workflows need different prompts for each stage. Managing state transitions and ensuring each step has the right context is error-prone when scattered across code.

The Solution

Define all workflow states in a single .prompt file. Each state gets its own instruction block, and the response contract tells your application what state comes next.

workflow.prompt

init do
  @version: 1.0
  @major: 1

  params:
    @workflow_state: enum[collect_info, analyze, generate, review]
    @user_input: str
    @collected_data: list[str]

  fragments:
    {system_prompt}: static from: workflows/system

end init

case @workflow_state do
collect_info: # Step 1
You are collecting information from the user.
Ask clarifying questions to gather needed details.
Focus on: `@collected_data`

analyze: # Step 2
You are analyzing the collected information.
Review what has been gathered:
{for data in @collected_data}
- {data}
{end for}

Provide a summary and identify gaps.

generate: # Step 3
Based on the analysis, generate the requested output.
Reference the collected data throughout.

review: # Step 4
Review the generated output for accuracy.
Suggest improvements if needed.
end @workflow_state

@user_input

response do
  {
    "workflow_state": "@workflow_state",
    "next_state": "...",
    "data_needed": []
  }
end response

Key Benefits

Single file defines entire workflow
State transitions are explicit in response contract
Fragments compose common workflow elements