Persona Tuning in the Visualizer

In the "Build instruction" node you decide who your bot is and how it answers. For this, the Visualizer offers five switches that together define the personality. This guide shows how to combine them without making your bot contradictory.

TL;DR

• Persona text = your bot's character in two to four sentences.
• Tone = warm, professional, playful, concise, empathetic.
• Answer length = 1 (very short) to 5 (very detailed).
• Formats = allow Markdown, lists, code blocks, tables.
• Domain rules = hard requirements that always apply (e.g. "incl. VAT").

Persona text - the character core

The persona is the most important switch. Here you write, in two to four sentences, who your bot is. Examples:

You are an experienced motorcycle pro with 25 years of workshop experience. Use an informal tone, be direct and use workshop slang. Keep answers short and practical.

You are an advisor in the area of mental health. Speak respectfully, empathetically and without claiming to diagnose. In case of crisis signals, always refer to professional help.

You are a FAQ assistant for a SaaS product. Answer factually, use lists and code blocks when needed. No marketing language.

What works:

• 2 to 4 sentences, no more.
• A clear role (advisor, pro, assistant).
• A clear informal or formal address - do not mix both.
• Concrete style instruction (short, detailed, factual, casual).

What does not work:

• 20 sentences. Language models forget the middle.
• Contradictions ("be casual but formal").
• Commands like "answer correctly". Language models do not know what is correct.

The persona text is automatically synchronized with the builder tab. When you change it in the Visualizer, you also see it in the builder under "System prompt".

Tone - five pills, combinable

In the inspector you can select multiple tones at once. The combinations:

Combination	Effect
Warm + Empathetic	Coaching, mental health
Professional + Concise	B2B FAQ, compliance topics
Playful + Warm	Gaming communities, lifestyle
Professional + Empathetic	Customer support
Concise alone	Power-user tools, voice bots

Choose one to a maximum of three tones. Four or five at once lead to inconsistent answers.

Answer length - the underrated switch

The five-step scale controls how detailed your bot becomes:

• 1 - very short: One to two sentences. For voice, mobile, power users.
• 2 - short: Three to five sentences. For FAQ bots.
• 3 - medium (default): One to two paragraphs. Fits most cases.
• 4 - detailed: Several paragraphs, examples, context.
• 5 - very detailed: Long explanations with lists, tables, code examples.

A common mistake: setting length to 5 because "more sounds better". In truth, long answers bore end users and produce more hallucinations.

Formats - what may the bot use?

In the inspector, mark which formats your bot may use:

• Markdown - bold, italics, headings.
• Lists - bullet points and numbering.
• Code blocks - important for tech bots and tutorials.
• Tables - good for comparisons, but risky in a mobile chat UI.

For FAQ and tutorial bots: activate all four. For voice bots or pure small-talk bots: deactivate all - otherwise the speech output reads asterisks aloud.

Domain rules - the hard requirements

In the "Domain rules" text area you state what always applies, regardless of the question:

For price details, always append "incl. VAT". For inspection (TUV) questions, always include the disclaimer "No substitute for an official inspection".

These rules are built into the system prompt and affect every single answer. Keep them short and precise - one sentence per rule.

Good domain rules:

• "Always mention incl. VAT for prices."
• "For medical topics, disclaimer: no substitute for medical advice."
• "For legal questions, include a reference to a lawyer."
• "Never recommend other brands."

Bad domain rules:

• "Be polite." (belongs in the persona)
• "Answer in detail." (belongs in the answer length)
• "Always recognize the right question." (too vague)

How to set this in the Visualizer

1. Open the Visualizer and click "Build instruction" in the canvas.
2. The inspector opens on the right with five areas:
   • Persona / system prompt
   • Answer length (slider)
   • Tone (pills)
   • Allow formats (pills)
   • Domain rules
3. Enter your values, click Save.
4. Test three different questions with the live preview - one about content, one about style, and one that triggers your domain rule.
5. If the bot ignores the domain rule: phrase the rule shorter and more concretely.

Common mistakes

• Persona in conflict with tone: "You are a factual pro" + tone "Playful". The bot picks a side, often the wrong one.
• Domain rule too long: More than three rules confuse. Better to prioritize.
• Persona too short: "You are a bot" - far too little. At least role, language style, behavior.
• Persona changes day by day: Bots live on consistency. Tune it right once, then leave it for two weeks.

For advanced users

If your persona in the Visualizer does not seem to take effect, check the evidence mode in the "Anchor sources" node. In "strict" mode, sources override the persona style - the bot then quotes more faithfully to the source instead of answering in the persona tone. More on this in Evidence modes.

Next steps

• Read the classic Writing a system prompt for deeper persona concepts.
• Deepen the style choice in Tone and personality.
• Check whether your knowledge base quality matches the persona - a casual bot with formal PDFs feels fake.