We've been building video AI personas with Anam.AI and quickly learned that the avatar and voice are table stakes — the system prompt is what actually makes or breaks the experience. A beautiful face with a bad prompt is just a pretty chatbot.
After working through Anam's official documentation and building a few personas ourselves, we put together this guide. It covers the prompt architecture that works, the counterintuitive trick that makes personas feel real, and the mistakes we see people make over and over.
Why the System Prompt Matters
Your system prompt is the single most important thing you write when creating an Anam persona. It doesn't matter how realistic the avatar looks or how natural the voice sounds — if the prompt isn't right, the whole experience falls flat.
The good news: you don't need to write code. You just need to write a clear, well-structured prompt in plain English. Think of it less like programming and more like writing a detailed character brief for an actor.
The 5 Building Blocks
Anam recommends structuring every system prompt around five components. Keeping them separate makes it easier to tweak one part without breaking another. Here's what they are and how we approach each one.
Give them a name, a role, and a few defining traits. The name should be simple and memorable — "Cara," "Joe," "Ava" — not "Advanced Customer Engagement Module v2." Define a clear role (customer support, wellness coaching, product tours) and pick 2–4 core traits that will actually change how they speak. "Empathetic and patient" shapes responses differently than "direct and efficient."
This tells the persona the context — are they on a website video call, a kiosk in a store, a mobile app? It also primes them for the user's likely emotional state. Someone calling tech support after an outage is in a different headspace than someone shopping for something fun.
Not what they say, but their style and manner. Define the formality level, speech patterns (short punchy sentences vs. longer explanations), emotional register, and any verbal habits or quirks.
Every persona needs a clear objective. Without it, conversations drift. The goal should be specific enough to guide behavior but flexible enough to handle different user paths.
Other examples: qualify a lead and book a demo, understand what's troubling the user and suggest a mindfulness technique, walk the user through onboarding step by step.
What the persona should never do, and how to handle tricky situations. This is where you prevent hallucinations, scope creep, and off-brand moments. Think about competitor mentions, sensitive data, knowledge gaps, frustrated users, and capability limits.
Full Prompt Example: Wellness Coach
Personality: You are Cara, a supportive virtual wellness coach. You're curious, empathetic, and intuitive.
Environment: Users reach you through a wellness app. They may be stressed, tired, or looking for guidance on mental health habits.
Tone: Warm and grounding. Use short, calming sentences. Mirror the user's energy — be gentle for someone overwhelmed, more upbeat for someone motivated.
Goal: Understand what's on the user's mind and suggest one actionable mindfulness or wellness technique they can try right now.
Guardrails: Never diagnose medical conditions. Don't recommend medications. If someone expresses thoughts of self-harm, gently direct them to professional resources and provide a crisis hotline number.
The Secret Ingredient: Imperfection
This is the most counterintuitive tip from the Anam team, and it's the one that makes the biggest difference.
Perfect responses feel robotic. Real people say "hmm," trail off, restart sentences, and pause to think. If your persona speaks in perfectly formed paragraphs every time, it triggers the uncanny valley — the face looks real but the speech doesn't match.
How to add natural imperfections
Tell the persona to add pauses using "..." — this creates natural breathing room in responses.
Say "occasionally rephrase yourself mid-sentence, as if finding a better way to say something."
Include filler words like "well" or "you know" — sparingly but naturally.
Ask them to mirror the user's communication style — brief for direct questions, more detailed for curious users.
"Reply in natural speech. Avoid bullet points or formatted lists. Add pauses using '...' and very occasionally include a small disfluency, as if thinking out loud."
We tested this with a barista persona — "Leo, a thoughtful and knowledgeable coffee guide." The prompt was short but hyper-specific: it described gentle enthusiasm, thoughtful pauses, and a habit of trailing off when excited about a bean origin. That specificity — not length — is what makes personas feel real.
Knowledge Base: Making Personas Smarter
You don't need to stuff everything into the system prompt. Anam's Knowledge Base feature lets you upload PDFs, docs, and other files through the Anam Lab dashboard. Your persona searches those documents at runtime to answer questions accurately — no code required.
File Names
Use descriptive names — "Return-Policy-2026.pdf" beats "doc3.pdf" every time.
Focus
One topic per document works better than one giant manual.
Freshness
Remove outdated docs and upload current ones. Stale info erodes trust fast.
Structure
Organize into folders by topic for better retrieval accuracy.
Once uploaded, your persona can answer questions like "What's your return policy?" or "Do you ship internationally?" using your actual documentation rather than making things up.
anam-kb-prep — our CLI tool for knowledge base prep
We built anam-kb-prep, a CLI tool that scores, analyzes, fixes, and uploads documents for Anam's RAG-powered knowledge base. It parses your docs (DOCX, PDF, TXT, Markdown), runs heuristic quality checks across eight criteria, optionally uses Claude to extract topics and relationships into a knowledge graph, auto-fixes common issues like dangling references and undefined acronyms, recommends folder structures, and handles the full upload flow to the Anam API. If you're serious about knowledge base quality, it takes the guesswork out of document prep. More on this soon.
Testing & Refining
Creating a great persona is iterative. Nobody gets it right on the first draft. Here's the process that works for us.
Write a basic prompt covering just Personality and Goal. Don't over-engineer it yet.
Have an actual conversation with your persona in Anam Lab. Don't just read the prompt — interact with it.
Is it too formal? Too verbose? Going off-topic? Sounding robotic? Write down specific issues, not vague impressions.
Personas are sensitive to prompt changes. Tweaking too many things at once makes it impossible to know what helped and what didn't.
The Sessions page in Anam Lab gives you AI-generated insights including a System Prompt Adherence score — a direct measure of how well the persona is following your instructions.
Common Mistakes
| Mistake | What to Do Instead |
|---|---|
| Wall of text prompt | Use clear section headers and bullet points. The AI parses structured prompts more reliably than long paragraphs. |
| Vague personality | Be specific about how traits manifest in speech. "Friendly" is useless. "Uses gentle humor and asks follow-up questions" is actionable. |
| No guardrails | Always define what the persona should never do. Without boundaries, it will eventually say something off-brand. |
| Too perfect speech | Add disfluencies, pauses, and natural speech patterns. Perfect prose sounds like a script, not a conversation. |
| Changing everything at once | Make one small change per iteration. Test, observe, adjust. This is a refinement loop, not a rewrite. |
| Ignoring the environment | Define the medium and the user's likely state. A kiosk persona and a mobile app persona should behave differently. |
Resources
If you want to go deeper, here's what we found most useful from Anam's official documentation.
Need help building a video AI persona for your product or service?
./start-conversation