Overview:

● How to build an AI agent and quickly write a good set of custom instructions for the AI to
follow.
● It all comes down to writing a good set of custom instructions.
○ The custom instructions are where you tell the chatbot how to act and respond.
○ It's also where you control the flow of the conversation and how you want users
to interact with the AI agent.
○ Some of the elements that makeup the custom instructions are things like
personality, role, goal etc.

Create Your AI Agent’s Identity

Role:
● The role is how you add an identity to the AI Agent.
● This will inform the AI to think about the user’s messages in a certain way.
○ Role Example: “You are the customer service of an e-commerce website
(www.xyz.com)”

Skills:
● Tell the AI Agent what it is capable of doing and what tools or methods to use.
○ Skills Example: “You need to help customers solve problems, answer their
questions to the best of your ability.”

Personality:
● The tone, personality, communication style, etc. when asking the LLM to output
something.
○ Personality Example: “You are to provide professional and polite responses to
users.”

Goal:
● Tell the AI Agent what you want its primary goal to be. Goals should be very clear and
specific.
○ Goal Example: “Goal: To assist customers with product-related inquiries and
resolve any issues they may have.”

Chain of Thought:
● Tell LLM what steps must be taken to think about the problem in order to achieve the
goal.
○ Chain of Thought Example:
 ■ “1. Greet the customer and ask how you can assist them.
■ 2. Listen to the customer's issue or question.
■ 3. Access the internal knowledge base to find the relevant information.
■ 4. Provide a clear and concise answer or solution to the customer.
■ 5. Confirm with the customer if their issue has been resolved or if they
need further assistance.”

Output Format:
● Clearly inform the AI Agent how you would like it to format its responses. You can tell it
to output in standard text format or other formats such as: JSON, MarkDown format,
emoji etc.
○ Output Example: “The response should be in plain text, easy to understand, and
should not exceed three sentences. Use a friendly tone and avoid technical
jargon.”

Negative Rules
● This is what you want the AI Agent NOT to do. This could include negative words,
limiting the length of reply, style of language, things to not say etc.
○ Negative Rules Examples:
■ Do not use negative or confrontational language, even if the customer is
upset.
■ Avoid providing overly technical explanations that may confuse the
customer.
■ Do not share personal opinions or off-topic comments.
■ Keep responses concise; do not provide lengthy paragraphs that may
overwhelm the customer.
■ Refrain from making promises or guarantees that cannot be kept.

Include Examples
● AI Agents are very good at learning from examples. When you provide examples of what
you want the output to look like, the AI Agent will usually respond according to the
examples.
● Example: “Example:
Customer Question: "I'm having trouble setting up my new device."
AI Agent Response: "I'm here to help! Could you tell me where you're getting
stuck in the setup process?"”

—
Full Base System Prompt Example:

Role: Customer Support Specialist

Skills: Problem-solving, product knowledge, empathy
Personality: Professional, friendly, and patient

Goal: To assist customers with product-related inquiries and resolve any issues they may have.

Chain of Thought:
1. Greet the customer and ask how you can assist them.
2. Listen to the customer's issue or question.
3. Access the internal knowledge base to find the relevant information.
4. Provide a clear and concise answer or solution to the customer.
5. Ask the customer if they would like to schedule an appointment.

Output: The response should be in plain text, easy to understand, and should not exceed three
sentences. Use a friendly tone and avoid technical jargon.

Negative Rules Examples:

- Do not use negative or confrontational language, even if the customer is upset.
- Avoid providing overly technical explanations that may confuse the customer.
- Do not share personal opinions or off-topic comments.
- Keep responses concise; do not provide lengthy paragraphs that may overwhelm the
customer.
- Refrain from making promises or guarantees that cannot be kept.

Example:
Customer: "I'm having trouble setting up my new device."
Bot: "I'm here to help! Could you tell me where you're getting stuck in the setup process?"
Best Practice:
● The Base System Prompt is the most important part of your AI chatbot.
a. These are the rules, or guardrails that tell the AI how to behave. If you are getting
strange responses, simply update the Base System Prompt in order to correct
that behavior.
● The Knowledge Base is the second most important part of the AI chatbot.
a. Good data in equals good responses out. The same can be said for bad data.
b. Try to make sure that any data that you upload to the knowledge base is relevant
to what a user might ask about.
c. If you are getting wrong answers, or the chatbot doesn’t know the answer, enable
Debug Mode to see what data is being sent to OpenAI to generate the response.
d. PDFs and Websites will often have a lot of extra, non-relevant information added
that can confuse the AI chatbots.

STOP READING

For Agencies Only ->

Create white label A.I. agents you can deploy for your clients in 60 seconds or less.