Best Practices, Pitfalls & How We Do It at SiteGuru.ai
Prompt engineering isn’t just for techies. It’s for anyone building AI into their product, site, or workflow.
It’s how you turn "just a chatbot" into an actually useful assistant — one that sounds like your brand, understands your users, and knows when to say “I’m not sure.”
At SiteGuru.ai, we’ve helped teams launch assistants that live on websites, in CRMs, and even on WhatsApp and SMS.
Here’s what we’ve learned along the way — and how we power it ourselves.
❌ Common Mistakes We’ve All Made
“Be helpful” is not a strategy.
Trying to cover everything in one paragraph.
Forgetting tone, fallback behavior, or escalation logic.
Writing prompts like code instead of conversation.
🧠 The mindset shift: Prompt design is interface design — for language.
✅ What Works (Consistently)
Write like you're briefing a smart teammate.
Use structure:
- Who the assistant is
- What it should do
- How it should speak
- What it should do when unsure
Focus on clarity, not cleverness.
Avoid ambiguity — it breaks trust faster than errors.
Test with real users — then revise.
🔍 How SiteGuru.ai Handles Prompting
We don’t just paste a paragraph into the model and hope for the best.
We built a layered instruction system that combines:
- Your brand's voice
- Chat history + user context
- Smart intent recognition
- Guardrails for trust, tone, and relevance
It starts simple — then scales up.
🟢 The 2 Required Instruction Fields
- "Who your Guru is" — Define tone, skills, language (e.g., “A friendly, multilingual assistant trained in Sales & Support.”)
- "What it is tasked with" — Define what it should handle — and just as importantly, what it shouldn’t.
That’s enough to get a great assistant live in minutes — because behind the scenes, we inject your instructions into a comprehensive instructions and logic layer designed to reduce AI “creativity,” detect hallucinations, and coordinate a multi-model, multi-step workflow that delivers accurate, reliable answers.
⚙️ The 3 Optional Advanced Fields
- Prompt Analysis Notes — Guide how the AI interprets vague prompts
- Reference Document Evaluation Notes — Filter out content that's off-topic
- Response Generation Notes — Control phrasing, fallback style, or markdown rules
🧠 Examples: “Only include plain-text URLs — this Guru runs on WhatsApp & SMS.” or “Disregard API docs unless the query is about integrations.”
🔗 What About URLs?
Whether your AI assistant includes links — and which links — is controlled in the Settings panel, not buried in prompt text.
You can:
- ✅ Enforce inclusion of citations (always validated to include only URLs explicitly included in your references)
- ❌ Forbid URLs entirely
- 🎯 Let the assistant decide — neither forced nor forbidden. The platform will still validate that only approved URLs make it through.
This ensures consistent, compliant responses — without the AI hallucinating links or violating policy.
💡 Final Takeaways
- Prompting isn’t technical — it’s strategic
- Most great assistants start with just 2 clear inputs
- Test > Listen > Iterate — your users will teach you everything
- The best prompts build trust by knowing when to say “I don’t know” — and when to escalate
At SiteGuru.ai, we believe prompt design should be:
✔️ Fast ✔️ Flexible ✔️ Brand-aligned ✔️ Developer-optional
You can change prompts on the fly, add fallback behavior, validate links — all without redeploying anything.
It’s prompt engineering that respects both your user and your team.
🚀 Ready to Try It?
👉 Curious what kind of help your assistant could offer your site visitors — with just two fields?
Set one up now at SiteGuru.ai — or message us, and we’ll get started together.
