Rasa Writing Responses

Writing Responses in Rasa

Workflow

1. Identify what the assistant needs to say and in which flow step.
2. Review existing responses in the domain — reuse before creating new ones.
3. Name the response with the utter_ or utter_ask_ prefix (see "Naming conventions").
4. Write the response text with slot variables where needed (see "Slot variables").
5. Add variations if the response should not sound identical every time

(see "Response variations").

1. Add buttons, images, or custom payloads for rich content if needed.

(see "Rich responses").

1. Add conditions if the response should change based on slot values or channel

(see "Conditional and channel-specific variations").

1. Enable rephrasing if needed (see "Rephrasing" and the

rasa-rephrasing-responses skill).

1. Reference the response in a flow as an action: utter_* step.

File organization

Responses are defined under the responses key in domain YAML files. Before creating new responses, check the existing domain layout — the user's existing convention always takes precedence.

Domain files are commonly split by business process, so each flow's slots, responses, and actions live together (e.g. domain/transfer_money.yml).

domain/
├── transfer_money.yml  # slots + responses + actions for one flow
├── book_appointment.yml
└── shared.yml          # responses reused across multiple flows

Naming conventions

This section proposes default conventions. If the user's project already follows different conventions, match those instead.

• Every response name must start with utter_:

utter_greet, utter_ask_amount, utter_transfer_complete

• Use snake_case after the prefix
• Name after the intent or purpose: utter_ask_email, utter_confirm_booking
• For collect steps, name responses utter_ask_{slot_name} — Rasa resolves this

automatically (e.g. collect: account_type → utter_ask_account_type). See the rasa-building-flows skill for override and custom action details.

• Always include a default variation without conditions as a fallback

Response basics

A response is a named template under the responses key. Each name maps to a list of one or more variations. Responses starting with utter_ can be used directly as flow actions without listing them in the actions section.

responses:
  utter_greet:
    - text: "Hey! How can I help you?"
  utter_bye:
    - text: "See you later!"

Slot variables

Insert slot values into responses using curly brackets. If the slot is empty or does not exist, the variable is replaced with None.

Values can also be passed from a custom action via dispatcher.utter_message(response="utter_greet_user", user_name="Sara").

responses:
  utter_greet_user:
    - text: "Hello, {user_name}! How can I assist you today?"

  utter_order_status:
    - text: "Your order {order_id} is currently {delivery_status}."

Response variations

Multiple variations under the same response name make the assistant sound less repetitive — Rasa picks one randomly at runtime. Optionally assign an id to each variation so an external NLG server can identify which was selected.

responses:
  utter_greet:
    - id: "greet_1"                              # optional, for NLG server
      text: "Hello! How can I help you?"
    - id: "greet_2"
      text: "Hi there! What can I do for you?"
    - text: "Hey! How may I assist you today?"   # id is optional

Rich responses

Buttons

Buttons give users structured choices. Each button has a title (displayed text) and a payload (message sent when clicked). Buttons skip the pipeline and directly annotate the user message.

In CALM, use /SetSlots(slot_name=value) payloads to set slots directly. Multiple slots can be set in one payload. Slot names must not contain (, ), =, , and slot values must not contain ,, (, ).

responses:
  utter_ask_card_type:
    - text: "Which card would you like to use?"
      buttons:
        - title: "Credit"
          payload: "/SetSlots(card_type=credit)"
        - title: "Debit"
          payload: "/SetSlots(card_type=debit)"
        - title: "Premium Credit"
          payload: "/SetSlots(card_type=credit, tier=premium)"  # multiple slots

Images

responses:
  utter_cheer_up:
    - text: "Here is something to cheer you up:"
      image: "https://i.imgur.com/nGF1K8f.jpg"

Custom payloads

Send arbitrary JSON to the output channel via the custom key. Styling and rendering are handled by the frontend or messaging channel.

responses:
  utter_date_picker:
    - custom:
        blocks:
          - type: section
            text:
              text: "Pick a date:"
              type: mrkdwn
            accessory:
              type: datepicker
              initial_date: "2026-01-01"

Conditional and channel-specific variations

Variations can be filtered by slot conditions, channel, or both. When combined, Rasa selects in this priority order:

1. Conditional match + matching channel
2. Default (no condition) + matching channel
3. Conditional match + no channel
4. Default + no channel

Conditional variations

Select a variation based on slot values using the condition key with predicate expressions (same syntax as flow conditions). Always include a default variation without a condition.

responses:
  utter_greet:
    - condition: slots.prior_visits > 1            # returning user
      text: "Welcome back, {name}! How are you?"
    - condition: not slots.prior_visits             # first visit
      text: "Welcome! How can I help you today?"
    - text: "Hello! How can I assist you?"          # default fallback

Channel-specific variations

Use the channel key to tailor responses per output channel. Rasa prefers channel-specific variations for the active channel and falls back to variations without a channel key.

responses:
  utter_ask_game:
    - text: "Which game would you like to play on Slack?"
      channel: "slack"
    - text: "Which game would you like to play?"    # all other channels

Overriding default pattern responses

Rasa has built-in default responses for patterns (e.g. utter_can_do_something_else). Override them by defining a response with the same name in your domain:

responses:
  utter_can_do_something_else:
    - text: "Is there anything else I can assist you with?"

Rephrasing

Responses can be dynamically rephrased by the Contextual Response Rephraser. Enable it per response with metadata: rephrase: True, or disable with rephrase: False. See the rasa-rephrasing-responses skill for full details.

responses:
  utter_greet:
    - text: "Hey! How can I help you?"
      metadata:
        rephrase: True

Voice-specific properties

For voice assistants, control whether users can interrupt the assistant while a response is being spoken using allow_interruptions. Use allow_interruptions: false for critical information (balances, disclaimers, compliance content, emergency instructions).

responses:
  utter_terms_and_conditions:
    - text: "Please note the following terms and conditions..."
      allow_interruptions: false     # user must hear the full message

  utter_ask_preference:
    - text: "What would you like to do today?"
      allow_interruptions: true      # default

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: RasaHQ
• Source: RasaHQ/rasa-agent-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.