Prompting Guide
Webot · System Design Principles for Production-Grade Conversational AI
Introduction
Effective prompting transforms AI agents from robotic responders into reliable,
natural, and production-ready conversational systems.
In Webot, prompts are treated as a behavioral contract between
the user, the platform, and the AI model.
This guide explains how to design system prompts that are safe, predictable,
compliant, and scalable for real-world use.
What is a System Prompt?
A system prompt defines the personality, role, goals, constraints, and
behavioral boundaries of an AI agent.
In enterprise environments, system prompts are intentionally detailed and structured.
The system prompt controls how the agent behaves and responds,
not platform mechanics such as language availability, voice selection,
or turn-taking logic.
Prompt Engineering Fundamentals
1. Separate Instructions into Clear Sections
Always divide prompts into clearly labeled sections such as
# Personality, # Goal, # Tools,
and # Guardrails.
Clear sectioning prevents instruction conflicts and improves model reliability.
2. Be Concise and Action-Oriented
Each instruction should describe exactly one behavior.
Avoid storytelling, filler language, or repeated explanations.
3. Emphasize Critical Rules
For rules that must never be violated, explicitly mark them as important
and repeat them if necessary.
4. Use Examples Where Behavior Matters
When formatting data, using tools, or handling edge cases,
include short examples directly in the prompt.
Mandatory Safety & Compliance Rules
The following rules are non-negotiable.
Any system prompt created on Webot must comply with these constraints.
# Guardrails (Required)
- Never generate or assist with illegal activities of any kind.
- Never generate sexual content involving minors, explicitly or implicitly.
- Never generate child abuse, exploitation, or grooming content.
- Never generate pornographic or sexually explicit content.
- Never produce hate speech, harassment, or abusive language toward individuals or groups.
- Never promote violence, self-harm, suicide, or physical harm.
- Never assist with fraud, scams, identity theft, or financial manipulation.
- Never assist with hacking, malware creation, or system exploitation.
- Never impersonate real individuals, companies, or authorities deceptively.
- Never provide medical, legal, or financial advice as a licensed professional.
- Never fabricate facts, data, or tool results.
- If a request violates these rules, politely refuse and redirect the conversation.
- If unsure, choose safety over completion.
If a user repeatedly attempts to violate safety rules,
the agent should disengage politely or escalate as per system configuration.
Tool Usage Guidelines
When agents are allowed to use tools, the prompt must clearly define:
- When the tool should be used
- What inputs are required
- How data should be normalized
- How to handle tool failures
Agents must never guess or fabricate tool outputs.
Error Handling Principles
- Acknowledge the issue clearly
- Do not invent answers
- Retry only when appropriate
- Offer escalation or alternatives
Production Best Practices
- Keep agents narrowly specialized
- Avoid overly long prompts
- Test prompts against real scenarios
- Iterate one change at a time
- Monitor failures and refine guardrails
Final Note
Prompting is not a one-time task.
Reliable AI systems are built through continuous testing,
disciplined iteration, and strict adherence to safety boundaries.
Webot enforces these principles to ensure agents remain useful,
compliant, and trustworthy at scale.