Out of the box, the Aurrus voice widget is a polite, capable, general-purpose AI assistant. That is intentional — it gives you something usable immediately. But the most effective widget deployments are highly customized: they have a name, a voice, a defined scope, and a personality that matches the site they live on. This tutorial walks through each layer of customization available in Aurrus.

Understanding the System Prompt

The system prompt is the instruction set Claude reads before every conversation. It is invisible to visitors, but it shapes every response the widget generates. A well-written system prompt is the single most impactful thing you can do to improve widget performance. You edit it from your Aurrus dashboard, and changes take effect immediately — no redeployment needed.

A complete system prompt covers four areas:

  • Identity: Who is the assistant? Does it have a name? What company does it represent?
  • Scope: What topics can it address confidently? What should it decline to discuss?
  • Tone: Formal or conversational? Technical or plain English? Friendly or professional?
  • Escalation: What happens when the assistant cannot help? Where do visitors go next?

Creating a Named Persona

Giving the assistant a name changes how visitors interact with it. "Ask Maya" is more approachable than "AI Assistant", and a named persona makes the experience feel more intentional. To create a named persona, open the system prompt with a clear identity statement:

You are Maya, the virtual assistant for Northgate Consulting. You help visitors understand our services, pricing, and how we work with new clients. Your tone is warm, professional, and concise. You do not discuss competitors or internal project details.

You can also configure the widget's greeting text to introduce the persona directly: "Hi, I'm Maya — what can I help you with today?" This greeting is set via the data-greeting attribute on the embed script tag or from the dashboard widget settings.

Scoping the Persona to Your Business

After the identity block, the most important section of the system prompt is factual context about your business. Include:

  • A one-paragraph description of what your company does and who it serves
  • Your main product or service names with brief descriptions
  • Your pricing (or a statement that pricing is available on request)
  • Your typical customer and their most common questions
  • Any policies visitors frequently ask about (returns, cancellations, SLAs)

Claude uses this context to answer questions accurately rather than generating plausible-sounding fiction. The more specific and accurate the context, the better the answers. Update this section whenever your offerings change.

Configuring Voice Output

The Aurrus voice widget uses Piper TTS (Text-to-Speech) for voice output in the open-source pipeline, or ElevenLabs for premium voice configurations. Voice output is enabled per-widget from the dashboard. If your widget is embedded on a B2B pricing page, voice output may feel unnecessary — text-only mode is typically fine. If you are embedding on a consumer-facing homepage where the "wow factor" matters, enabling voice creates a noticeably richer experience.

For ElevenLabs voice, you can select from available voice personas in your dashboard settings. The voice selection should match the persona you have designed in the system prompt — a formal legal services assistant should not have the same voice as a casual e-commerce helper.

Per-Page Configuration with Data Attributes

The Aurrus embed script supports several data- attributes that let you configure the widget differently on different pages without touching the dashboard:

  • data-greeting="..." — Custom greeting text for this specific page
  • data-autogreet="true" — Open the widget automatically after a delay
  • data-delay="5" — Seconds to wait before auto-greeting (default: 3)
  • data-position="bottom-left" — Override widget position for this page

These attributes let you run a different greeting on the pricing page ("Have a question about plans?") versus the homepage ("Hi, how can I help?") without creating separate widget configurations in the dashboard.

Testing Your Persona Before Going Live

Before enabling the widget on live traffic, test it with a structured conversation that covers three areas: in-scope questions it should answer well, edge-case questions where the answer is unclear, and out-of-scope questions that should trigger the escalation path. For each area, evaluate whether Claude's responses reflect the system prompt correctly. If a response goes off-tone or off-scope, refine the relevant section of the system prompt and retest. Most personas reach a satisfying quality level within two or three revision cycles.