Skip to content
AskVault Docs
Try Free →

Prompt engineering for AI customer support

Written by Joseph Russo, Principal Architect, Networks and Tech, AskVault · Reviewed by Allison Chan
Last updated: · 6 min read

What a system prompt is

A system prompt is the instructions the LLM reads before every customer message. It tells the model what role to play, what tone to use, what to refuse, what to cite, and how to handle uncertainty.

AskVault's default system prompt is built for customer support. Out of the box it produces grounded, source-cited, polite responses that decline rather than invent facts. Many teams can start with the defaults.

Customize it when you need a particular brand voice, industry vocabulary or specific refusal wording.

What you can customize

Under Settings > AI Configuration there is one place to shape behavior, plus one dial:

  1. Custom instructions (the system prompt). Plain-English text in the Structured view (Identity, Instructions, Constraints) or the Raw view (one text box). Tone, persona and refusal wording all live here.
  2. Temperature. A slider from Precise to Creative. A low value gives consistent, factual answers. The default is 0.3.

There are no separate tone presets, persona fields or refusal-message forms. If you want a formal tone, a named persona or a specific way of declining, write it into the instructions.

Underneath, AskVault's core instructions (answer from your content, cite sources, follow policy bounds) stay locked. They are load-bearing for accuracy; don't try to write around them.

Tone

Describe the tone in a sentence or two in your instructions. Three examples you can adapt:

Formal.

Use formal English. Avoid contractions. Address the customer as "you" but maintain professional distance. Cite sources for every claim.

Use for: legal services, financial services, healthcare, government.

Friendly-professional.

Use natural English with contractions where appropriate. Be warm but stay focused on the question. Cite sources for substantive claims.

Use for: B2B SaaS, e-commerce, most consumer-facing services.

Casual.

Use conversational English. Contractions, occasional slang, exclamation points are fine. Match the customer's energy. Cite sources but lightly.

Use for: gaming, lifestyle, younger-demographic consumer brands.

These are starting points to paste into the instructions and edit, not built-in presets.

Persona and AI disclosure

The persona is how the bot introduces itself. Write it into the Identity part of your instructions, for example "You are Sarah, the support assistant for Acme. If asked, say you are an AI assistant."

Some brands prefer a persona that never says it is an AI. That is risky in some places. The EU AI Act and several US state laws can require you to tell people they are talking to an automated system. For most B2B SaaS, saying so plainly is the safer default, and customers adjust their expectations accordingly. Check the rules that apply to you with your own legal advisor.

Strict mode

Strict mode is the most important safety behavior, and it is on by default for every workspace:

  • Strict (default). The bot answers only from your retrieved knowledge. If retrieval is weak, it declines and offers escalation.
  • Helpful (looser). The bot may combine retrieved knowledge with general knowledge. It answers more questions, including ones with weak retrieval, but the risk of a made-up answer goes up.

A clear "I don't know" does less damage than a wrong answer. Strict mode reduces made-up answers; it does not eliminate them, so keep checking real conversations. You can review the setting under Knowledge > Settings > AI Search Behavior, and the Playground has a Strict toggle so you can see how a question behaves in each mode before you change anything live.

Refusal wording

When the bot can't answer, what does it say? Write the wording you want into your instructions. Examples:

  • Out of scope. "That's outside what I can help with from our documentation. Want me to connect you with a human?"
  • Knowledge gap. "I don't have specific information about that in my knowledge base. Let me get a human to look at this."
  • Sensitive topic. "I'm not able to discuss that here. Let me connect you with the right team."
  • Policy refusal. "I can't help with that based on our policy. Here's how to reach a human agent."

There is no separate Refusal Messages screen. Default phrasing is neutral and offers escalation.

What not to change

Three parts of AskVault's system prompt stay locked:

  1. The "use only retrieved context" instruction. Loosening this re-introduces hallucination. The core RAG safety guarantee depends on it.
  2. The "cite sources" instruction. Customers expect citations; loosening it breaks trust.
  3. Policy bounds on skills. Skills like discount_negotiator have hard caps. Don't try to override them in the system prompt; they're enforced at the policy layer below the LLM.

Channels

One set of instructions applies to every channel. There are no per-channel prompt overrides. If a channel needs a different style (short, no markdown for voice; lighter for WhatsApp), say so in the instructions, for example "On voice, keep answers to two sentences and avoid lists."

Testing prompt changes

Changes go through a draft first:

  1. Edit your instructions and save the draft. Deployed channels keep using the live version.
  2. Open the Playground, which answers with your draft, and try real customer questions. Include the awkward ones.
  3. When you are satisfied, publish to push it to all live channels.

If the draft is not right, discard it.

Common pitfalls

Prompt too aggressive. "Never refuse, always answer." Forces hallucination. Keep strict mode on for B2B support.

Persona inconsistency. Conflicting personas in your instructions confuse customers. Pick one persona and write it once.

Refusal too apologetic. "I'm so sorry I cannot help with this" loops 5 times per conversation. Tone down. One clean refusal is enough.

System prompt versioning. A change broke retrieval for a specific query type. AskVault doesn't keep a version history to step back through: the draft/publish split means an unpublished change is easy to discard, but once it's live, undoing it means editing the prompt back by hand (or using Reset to start over from the default).

FAQ

Can I see what system prompt is in effect?

Partly. Settings > AI Configuration shows the system prompt you can edit, in Structured or Raw view. AskVault adds its own core instructions on top of it, and those are not shown or editable.

Does the system prompt affect cost?

Marginally. A longer prompt uses a few more tokens on every query. Keep your instructions short and specific.

Can I include examples (few-shot) in the prompt?

There's no dedicated examples field. Fold a couple of representative question-and-answer pairs into the custom instructions text itself if you need the model to match a very specific output format.

How do I test edge cases?

Keep a short list of representative customer questions, including the tricky ones. Run them in the Playground against your draft before you publish, and run them again after any change.

What if my industry needs a non-default safety stance?

Healthcare, legal, financial. Write the disclaimers and refusal wording you need into your custom instructions, test them in the Playground, then publish. AskVault does not configure a per-customer industry template, so have your own compliance team review the wording.

Was this page helpful?