Skip to main content
0 reactions

Rasa Tutorial — Part 7: Responses: Control the Exact Wording

Rasa Tutorial Part 7 card titled Responses: Control the Exact Wording. A responses.yml entry named utter_book_appointment_confirmed with the text “Your appointment is booked.” points to the agent’s locked reply “Your appointment is booked. Your booking reference is APT-7860.”, with the notes “Sent word for word, every time” and “Your own wrap-up question”.
Rasa Tutorial Part 7: Responses: Control the Exact Wording

Some sentences have to come out the same every time: a booking reference, a legal notice, your company’s own way of closing a conversation. In this part of the Rasa Tutorial, you give your Rasa Mantle agent responses: messages you write once in a responses.yml file, which Rasa sends exactly as written, without Claude rewording them.

By the end, the banking agent from Part 6 will confirm every appointment with the same sentence and the exact booking reference, and finish each task with your own wrap-up question instead of Rasa’s. You’ll see which messages came from your files in the Rasa Inspector.

This is Part 7 of the Rasa Tutorial. The parts in Module B, Building Skills:

All parts of the Rasa Tutorial

Module A: Start Here

Module B: Building Skills

Module C: Progressive Control

  • Part 8: Tool Constraints
  • Part 9: Scoped Instructions
  • Part 10: Ordered Blocks
  • Part 11: Sub-skills and Composition

Module D: Orchestration

  • Part 12: The Runtime Loop: How Mantle Handles a Turn
  • Part 13: Guarantees and Guardrails
  • Part 14: Skills and Routing
  • Part 15: Context Management

Module E: Agent Setup

  • Part 16: agent.yml: Identity, Persona and Rules
  • Part 17: integrations.yml: Claude, Model Groups and Routing
  • Part 18: MCP Tools
  • Part 19: Hooks
  • Part 20: Build Faster with the Rasa Copilot

Module F: Voice

  • Part 21: Talk to Your Agent in the Browser
  • Part 22: Designing for Voice
  • Part 23: Connect a Phone Line
  • Part 24: Monitor Voice Calls

Module G: Channels

  • Part 25: The REST Channel and a Website Chat Widget
  • Part 26: Microsoft Teams

Module H: Observability

  • Part 27: Tracing with Langfuse
  • Part 28: Metrics and Conversation Events

Module I: Deploy

  • Part 29: Run Your Agent with Docker
  • Part 30: CI/CD for a Rasa Agent
  • Part 31: Deploy to Kubernetes with Rasa’s Helm Chart
  • Part 32: Deploy on AWS
  • Part 33: Deploy on Azure
  • Part 34: Deploy on Google Cloud

Module J: Rasa Studio

  • Part 35: Rasa Studio with a Mantle Agent

Module K: Wrap-up

  • Part 36: Final Project, Cheat Sheet and What to Learn Next

Cost: Rasa’s Developer Edition licence is free. Claude is pay-per-use: Claude Sonnet 5.5, the model this series uses, costs $2 per million input tokens and $10 per million output tokens (Claude pricing (opens in a new tab)). In this part you send two messages, which make four calls to Claude: about $0.02. That’s an estimate: in Part 2, eight calls cost $0.05 on the Claude Console. Training also indexes Part 6’s reference document with Azure OpenAI again, which costs far less than a cent. The responses themselves cost nothing: Rasa sends them without calling Claude.

Prerequisites

Step 1: Write a Response

Open Terminal, go to your project folder and turn on its virtual environment:

TerminalBash
cd my-rasa-agent
source .venv/bin/activate

A response is a message you write word for word. Rasa sends it exactly as written, and Claude never rewords it. Use one when the exact words matter: a legal or compliance notice, a value the customer must get exactly right, or wording your brand requires. Leave everything else to Claude and your skill’s instructions, so the agent still sounds natural.

In Part 5, Claude wrote the booking confirmation itself, in its own words. Make it exact: in your text editor, create a file named responses.yml in the skills/book_appointment folder:

skills/book_appointment/responses.ymlYAML
responses:
  utter_book_appointment_confirmed:
    - text: >-
        Your appointment is booked. Your booking reference is
        {booking_reference}. Please bring a photo ID with you.
  • responses holds every response in the file.
  • utter_book_appointment_confirmed is the response’s name. All response names in a project share one list, so Rasa’s docs advise naming each one after its skill. Rasa’s own responses start with utter_ too.
  • Under the name is a list of versions, each starting with - text. This response has one version, so that’s the one Rasa sends.
  • >- lets the text run over several lines in the file. YAML joins them into one line.
  • {booking_reference} is filled in from the skill’s memory when the message is sent: the reference that the book_appointment tool from Part 5 saves.

Step 2: Send the Response When the Booking Is Made

Rasa sends a response only when something asks for it. Open skills/book_appointment/skill.md and add an utter list to the frontmatter, after description:

skills/book_appointment/skill.md (excerpt)YAML
utter:
  - utter_book_appointment_confirmed:
      when: "session.book_appointment.booking_reference"
  • utter lists the responses this skill sends by itself.
  • when sends the response the first time its condition becomes true. A memory field on its own, written as session.<skill folder>.<field>, means “this field has a value”. So the confirmation goes out as soon as the tool saves a booking reference.
  • To send a response the moment a skill starts instead, write on: activate in place of when.

The confirmation now gives the reference, so take that job away from Claude. Change step 3 of the instructions to:

skills/book_appointment/skill.md (excerpt)Plain text
3. When they confirm, book it with @tool.book_appointment.

Rasa also tells Claude not to repeat a response it has just sent. Here’s the whole file now:

skills/book_appointment/skill.mdPlain text
---
name: Book an Appointment
description: Branch appointments - book a meeting with an adviser at one of our branches

utter:
  - utter_book_appointment_confirmed:
      when: "session.book_appointment.booking_reference"
---

Help the customer book an appointment with an adviser at a branch.

1. Find out which branch they want to visit, what the appointment is about
   and which day suits them. Ask for anything they haven't said, one question
   at a time. Record each answer with `set_memory` as soon as they give it.
2. When you have all three, read them back and ask the customer to confirm.
3. When they confirm, book it with @tool.book_appointment.

Advertisement

Step 3: Replace Rasa’s Wrap-Up Question

When a task finishes, Rasa’s built-in Completed skill asks whether the customer needs anything else. Its words come from a response named utter_ask_wants_to_continue: “Is there anything else I can help you with?” Rasa ships its built-in skills’ responses with rephrase: true, so Claude may reword this question each time, in a separate call.

To use your own words, add a response with the same name to a responses.yml file in your project folder, next to agent.yml:

responses.ymlYAML
responses:
  utter_ask_wants_to_continue:
    - text: >-
        Is there anything else I can help you with today, such as
        finding a branch or booking an appointment?
  • Rasa merges every responses.yml into one list and reads your project folder’s file last, so your utter_ask_wants_to_continue replaces Rasa’s.
  • Without rephrase: true, Rasa sends your text exactly as written, with no call to Claude.
  • Rasa’s other built-in messages, such as the satisfaction question that follows when the customer needs nothing else, can be replaced the same way. Their names are on Rasa’s Builtin skills (opens in a new tab) page.

Step 4: Train Your Agent

TerminalBash
rasa train

Rasa checks your responses while it trains: a response name in skill.md that isn’t in any responses.yml stops the training with an error, before the agent ever runs. When it finishes, the last line names your new agent package:

Output (the line that matters)Plain text
2026-11-21 16:33:57 INFO     rasa.mantle.model_archive.snapshot  - [info     ] Mantle agent package saved at 'models/20261121-163354-small-badminton.tar.gz'.

The INFO notes above it are the same as in Part 6; they aren’t errors.

Step 5: Book an Appointment

TerminalBash
rasa inspect

When the Rasa Inspector opens in your browser, send I'd like to book an appointment at Riverton Station to talk about a loan on Saturday. Claude reads the details back and asks you to confirm. Reply Yes, please.

A conversation in the Rasa Inspector: the customer asks to book an appointment at Riverton Station about a loan on Saturday, the agent reads the details back, the customer replies “Yes, please.”, and the agent answers with the exact text “Your appointment is booked. Your booking reference is APT-7860. Please bring a photo ID with you.”, followed by “Is there anything else I can help you with today, such as finding a branch or booking an appointment?”
The confirmation and the wrap-up question, word for word from responses.yml.

The confirmation is your text, word for word, with the booking reference filled in. Yours will show a different reference: the demo tool makes up a new one for each booking. The question after it is your wrap-up question, not Rasa’s.

Click the Inspector mode button at the top of the chat and choose Inspect. Under Yes, please., every message says where it came from:

Inspect mode in the Rasa Inspector under the message “Yes, please.”: an LLM call to claude-sonnet-5-5, Tool call: book_appointment, the booking confirmation marked utter_book_appointment_confirmed template, a second LLM call, the Book an Appointment skill completed, the Completed skill activated, Tool call: complete, and the wrap-up question marked utter_ask_wants_to_continue template.
Each response is marked with its name and “template”.
  • utter_book_appointment_confirmed template comes right after Tool call: book_appointment: the tool saved the booking reference, and your when condition sent the response.
  • A message Claude writes itself is marked LLM generated instead, like the read-back of your booking details.
  • utter_ask_wants_to_continue template is your wrap-up question. No LLM call comes before it: Rasa sent it as written.

When you’re done, go back to Terminal and press Ctrl+C to stop the agent.

Clean Up

If the agent is still running, press Ctrl+C in Terminal to stop it. Claude and Azure OpenAI both charge only for what your agent uses, so a stopped agent costs nothing.

Keep the my-rasa-agent folder: Part 8 builds on it.

Common Mistakes

MistakeDo this instead
Naming the file response.yml, or saving it outside the skill’s folder and the project folderName it responses.yml, and save it in the skill’s folder or next to agent.yml.
Giving two responses the same nameName each response after its skill. All names share one list, so a second response with the same name replaces the first.
Sending a response before its placeholder has a valueUse when on the field the placeholder reads. A placeholder with no value is sent as written, braces and all.
Asking Claude in the instructions to say what a response already saysLet the response say it, and keep the instructions to the steps Claude takes.
Adding rephrase: true to wording that must be exactLeave it out for legal, compliance and reference-number messages.

Troubleshooting

ProblemFix
rasa train stops with an error about a response nameMake sure the name in skill.md matches the name in responses.yml exactly, and that the file starts with responses:.
The chat shows {booking_reference} with its bracesThe field had no value when the response was sent. Check that the placeholder matches the field’s name in memory.yml, and send the response with when on that field.
The confirmation never appearsCheck that the when condition names the skill’s folder and the field, session.book_appointment.booking_reference. Then train again and restart rasa inspect.
The agent still asks Rasa’s own wrap-up questionSave your responses.yml next to agent.yml and spell the name utter_ask_wants_to_continue exactly. Then train again and restart rasa inspect.

FAQ

Can a response have more than one wording?

Yes. Add more versions under its name, each with a condition, plus one version without a condition as the fallback. Rasa sends the first version whose condition is true, such as different fee wording for a lost card and a stolen one.

Can Claude still adapt a response to the conversation?

Yes. Add metadata: with rephrase: true under the version’s text. Claude then rewords it in a separate call that has no tools, and Rasa sends your text as written if that call fails. Never do this for wording that must be exact.

Can I make my agent greet customers first?

Not with a response. Rasa’s built-in Session Start skill doesn’t speak, so it has no greeting to replace. To greet first, you replace that skill itself, as Rasa’s Start a conversation (opens in a new tab) guide shows.

Should every message be a response?

No. Rasa’s docs warn that an agent made only of responses sounds like a scripted phone tree. Use responses where the exact words matter, and let Claude phrase the rest from your skill’s instructions.

Conclusion

You gave your agent responses: exact wording in responses.yml, sent by an utter trigger the moment the booking reference is saved, and your own wrap-up question in place of Rasa’s. In Inspect mode, every response is marked as a template, so you can always tell your words from Claude’s.

← Previous: Part 6: References: Answer from Your Documents

Coming next: Part 8: Tool Constraints.

Additional Resources

Rasa, Claude, Azure and other names mentioned are trademarks of their respective owners. JD Bots isn’t affiliated with or endorsed by Rasa Technologies, Anthropic or Microsoft.

0 comments