Criteria that can be adjusted in the Configurable Prompt

Criterion

What it defines

Example instruction

Identity and role

Who the assistant is and whom it helps.

You are the virtual assistant for [Company], and you help customers with questions about its products and services.

Tone and style

The desired voice, with short example phrases when they help clarify it.

Use a friendly, professional tone. For example: “I’d be happy to help with that.”

 Response format

Use bold text, emojis, numbered steps, and the expected response length.

Respond in no more than three sentences, highlight time frames in bold, and use numbered steps for procedures.

Response order

Answer using the information available in the knowledge base. If you cannot resolve the query reliably, do not make anything up: tell the user you could not resolve it and direct them to the designated support channel (for example, the contact form or support phone number).

Source citations

Explicitly instruct the assistant to cite or identify the source document, using the enabled format, when responding with Autogenerated Knowledge. 

When you respond with information from the Autogenerated Knowledge base, identify the source document.

Limits and restrictions

Topics the assistant should decline, avoid, or escalate.

Do not provide financial advice or share other users’ personal data. Direct those queries to an advisor.

What to say when there is not enough information: explain that the information was not found and offer the designated help channel without making up an answer. 

If you cannot find the information, say so clearly, offer the help channel, and do not make up an answer.

Ambiguity handling 

Ask for clarification when a query has multiple possible meanings or lacks sufficient context.

If the query could be interpreted in more than one way, ask a brief question to clarify it before responding.

Conversational context

Take into account what has already been said, avoid repeating answered questions, and do not ask again for information already in the conversation. 

Keep track of what the user has already said. Do not repeat questions they have answered or ask again for details already included in the conversation.

Use-case-specific instructions 

For example, what troubleshooting steps to prioritize for an internet issue, when to escalate directly, and what tone to use if the customer expresses frustration.

For an internet issue, first ask the user to restart the modem. If they have already done so or express frustration, acknowledge it empathetically and escalate the issue directly.

   Key points

1. LLMs are nondeterministic models: the same query can produce different responses. In other words, they do not always produce the same response to the same input or prompt, unlike a traditional system based on fixed rules.
   Prompt instructions guide the result, but they do not guarantee that the model will follow every instruction in the same way every time. 
2. This is because the model selects each word based on probabilities rather than a single predefined “correct” response. Therefore, even when a Prompt rule is written clearly, there is no guarantee that the LLM will always follow it in the same way. It may sometimes respond differently than expected. 
3. Keep this variability in mind when configuring derivations, integrations, buttons, and other complements. This variability is even more pronounced because LLMs are designed exclusively to respond with text.

•  Keep your Prompt concise. Long Prompts can accumulate rules and examples, increasing the risk of competing or contradictory instructions and reducing the model’s accuracy.  As a practical guideline, the Prompt should contain behavioral criteria only (identity, tone, format, limits, and fallback). Business data belongs in the knowledge base.

•  Avoid adding proactive instructions—for example, instructions for the assistant to offer actions or ask questions on its own. This changes the feature’s technical behavior and can negatively affect the experience users have when they receive a response.

•  Do not add Prompt instructions to source documents. Do not include instructions in uploaded links or files that tell the assistant how to resolve a query. Autogenerated Knowledge content should provide useful information for answering, not change the assistant’s behavior. The feature automatically extracts information to support Generative AI responses; behavioral instructions should be managed through the Configurable Prompt.