Flow designerAbout nodes

Configure step nodes

Instruction Node

An Instruction Node is your catch-all building block for any discrete “step” in a flow that isn’t purely a conditional fork or an API call. It’s where you tell the AI what to do or say, and optionally attach actions to execute. 

Think of it like a general-purpose container: you give it a name, write instructions, wire in any variables, or action results, and the flow moves on once those actions complete (or immediately if there are none).

1. Purpose

  • Primary Use: Encapsulate a single, self-contained step—this could be a spoken reply, a data lookup, or triggering a function.

  • Versatility: an Instruction Node can combine a conversational instruction and one or more actions (API calls, database writes, etc.).

  • Flow Control: the node simply executes the instruction and immediately hands off to its next node

Instruction Node Fields & Configuration

  1. Step Name

    • What It Is: The identifier for this step. The AI agent reads this first.

    • Best Practice: Use a short verb-driven phrase that describes this step’s role—e.g., “Gather Email,” “Fetch Recent Orders,” “Confirm Subscription,” “Send Follow-Up.”

    • Why It Matters: Clear step names make your flow easy to scan, debug, and maintain by you and also by the AI agent.

  1. Instructions

    • What It Is: A free-form text box where you describe exactly what you want the AI Assistant to do or say at this step.

    • Formatting Options:

      1. Variables {}: Insert any dynamic data placeholders (e.g., {user_name}, {order_id}). The flow replaces these with real values at runtime.

      2. Action Results <: Pull in results from previously executed actions (e.g., <getUserProfile.name>, <fetchWeather.forecast>).

    • Guidelines:

      1. Be Explicit: If you need the assistant to ask a question, say “Ask the user: ‘What is your email address?’ or ‘Ask the user email”

      2. Keep It Focused: Don’t cram multiple responsibilities into one Instruction. If you need first to greet, then to validate data, consider splitting them into separate steps.

  2. Variables

    • What They Are: Named data points that flow in or out of this step.

    • How They Work:

      1. Inputs: Variables you reference in the Instruction (e.g., {user_id}) must be defined here—or inherited from an earlier node.

      2. Outputs: If your Instruction or Custom Action sets a new variable (e.g., “Store the user’s answer in {email_address}”), declare it here so downstream nodes can read it.

    • Why It Matters: Properly declared variables ensure data flows predictably through your conversation. Undeclared/mismatched names cause runtime errors.

    4. Extract Variables During Execution

  • What It Is: A mechanism to pull out and persist specific pieces of information from the AI’s generated response at runtime.

  • How to Configure:

    1. Under Extract Variables, click Add Variable.

    2. Name: Give the variable a clear identifier (e.g. issue_type, confirmation_code).

      • ! Very important to have clear variable names,

    3. Type: Choose String, Number, Integer, Boolean, List, Object

    4. Extraction Rule: Provide a simple instruction that tells the system how to extract the user’s reply (e.g., “What is the user’s email?)

    5. Required: If checked, the flow will loop back or trigger an error branch when extraction fails.

  • Runtime Behavior: After the AI sends its message, the engine applies your extraction rule, stores the result in {your_variable}, and makes it available to all downstream nodes—just like a collected or action-result variable.

Use Extract Variables when you need to capture free-form data directly from the AI’s generated text without an explicit “Ask the user…” prompt.

Best Practices & Tips

  1. Name Consciously

    • Bad: “Step 7” or “Do Something.”

    • Good: “Validate Email Format,” “Charge Card,” “Present Upsell Offer.”

  2. Single Responsibility

    • Each Step Node should do one primary thing: either speak/ask, or set context. If you mix too many actions and a long prompt in one node, it becomes hard to manage.

  3. Avoid Long, Block-Style Instructions

    • If your Instruction reads like a short essay (five sentences or more), consider breaking it into two smaller steps: one to outline context, another to ask/act.

    • You or any human should be able to glance at “Instructions” and understand what the AI is doing.