> ## Documentation Index
> Fetch the complete documentation index at: https://docs.hubtalk.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# 04. Conversation

> A static phrase or a prompt, node-level overrides and Skip response.

# Conversation: static phrase or prompt

> [!NOTE]
>
> **Module goal:** learn to choose between a fixed phrase and an LLM instruction and understand the Node-level overrides block.

A speaking node has two forms of content — the **Prompt** and **Static** tabs in the **Reply** block — and they solve different tasks:

| Form | What it does | When to choose it |
| - | - | - |
| **Static** (`static`) | Spoken word for word on the first entry into the node, without calling the model; after that the model carries on "in its spirit". It sounds in the language it is written in: the model does not translate it. | A greeting, legal wording, any words that must not be paraphrased. A bonus: a phrase without `{{variables}}` is synthesized in advance, at publishing, and plays without delay; a phrase with variables gets into the synthesis cache during calls. |
| **Prompt** (`prompt`) | An instruction for the model: the reply is generated anew every time. | Where the answer depends on the conversation: clarify the reason, offer time slots, calm down an unhappy customer. |

`{{variables}}` work in both forms. The **Open in editor** button opens the reply full screen; on the right there is a **Variables** list, and a click inserts a variable at the cursor. This is the easiest way to insert the customer's name, the visit date or the request number without a typo in the variable name.

> [!TIP]
>
> **Tip: the phrase and the language of the conversation.** If the agent speaks two languages — for example, Kazakh and Russian (`kk`/`ru`) or English and Spanish (`en`/`es`) — a fixed phrase is still spoken in the language you wrote it in. Where the customer's language is not known in advance, choose **Prompt**; keep **Static** for texts that must sound exactly in this wording.

![Conversation node inspector: fixed phrase, variables, overrides, transitions](https://mintcdn.com/hubtalk/UuNkjJ26eB-xpPwv/images/tutorial/tut-07-node-conversation_en.webp?fit=max\&auto=format\&n=UuNkjJ26eB-xpPwv\&q=85\&s=45795e4ea7b90aa888c34f626c80dd2a)

The Visit confirmation node. The agent's first conversation node: a fixed phrase with four variables and four transitions by meaning — SMS reminder, rescheduling, call back later and refusal.

> [accordion] Show the whole panel
>
> ![Conversation node inspector: fixed phrase, variables, overrides, transitions](https://mintcdn.com/hubtalk/UuNkjJ26eB-xpPwv/images/tutorial/tut-07-node-conversation_en.webp?fit=max\&auto=format\&n=UuNkjJ26eB-xpPwv\&q=85\&s=45795e4ea7b90aa888c34f626c80dd2a)

## Node-level overrides

The **Node-level overrides** block sits in the collapsed **Advanced** section at the bottom of the inspector; when something is set, the section title shows how many settings are set, and the section is open right away. A switch that is off = the node inherits the agent's setting. A switch that is on applies only while this node is speaking. This is fine-tuning, not a global setting — it fixes specific spots in the dialog:

* **Interruption** — whether the node's reply can be interrupted and how easily. **Allow interruption** unticked ⇒ the node finishes its reply in full (needed for disclaimers). The **Interruption Sensitivity** slider: 0 — only a long phrase interrupts, 1 — the agent stops immediately. Zero does not turn interruption off: to make the node finish speaking, untick the switch.
* **Generation model for this node** — a different generation model for this node only, picked from the directory. Useful when one complex node needs a stronger model while the agent as a whole runs on a fast and cheap one. The transition decision out of this node is still made by the agent's classification model.
* **Voice** — a different TTS voice while this node is speaking, also from the directory.
* **Knowledge bases** — a separate row of the same **Advanced** section: its own set of bases instead of the agent's (see [module 09, Knowledge bases](https://docs.hubtalk.ai/v4/tutorial/knowledge-base.md)).

> [!NOTE]
>
> **Why Skip response changes the flow**
>
> The **Skip response** switch turns the node into "say it and move on": such a node has exactly one transition, and it becomes `Always`. This is how service lines before the next question are built — for example, "Just a reminder: your appointment is at 3 PM with Dr. Smith". The agent does not wait for the customer's answer and moves on right away. "Connecting you to the front desk" does not need a separate node: `Transfer` says this phrase itself (see [module 11, Transfer](https://docs.hubtalk.ai/v4/tutorial/transfer.md)).

> [!WARNING]
>
> **Pitfall: the node prompt ≠ the agent prompt**
>
> The agent's global prompt sets the personality and the prohibitions ("don't make up prices"), the node instruction sets the task of a single step. If a rule is needed throughout the conversation, it lives in the global prompt: otherwise you will duplicate it in twenty nodes and one day forget it in one of them.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.