AgentFlow lets you design smart, contextual AI agents using natural-language prompts. Whether you’re handling customer support, lead qualification, or FAQ automation, how you write your prompts directly impacts how your AI agent responds.
This article will guide you through best practices for writing effective prompts and conditions — from tone and behavior rules to fallback logic and exit criteria. You’ll also find industry-specific prompt examples to help you get started quickly, whether you’re in e-commerce, education, or other industries.
How prompts shape your AI agent’s behavior
Writing a prompt isn’t just about telling the AI what to say — it’s about shaping how your agent thinks, reacts, and responds across different situations.
In AgentFlow, your AI agent relies on prompts to understand three things:
- What it represents (your brand, team, or role)
- What it should and shouldn’t do (task scope, tone, guardrails)
- How to behave in real conversations (reply style, when to exit, what signals to look for)
If a prompt is too vague, your agent may give generic or unhelpful replies. If it’s overly detailed or unfocused, the agent may miss what matters most. The goal is to give just enough context so the agent knows how to act — without overwhelming it.
What makes a good prompt?
A well-written prompt is what turns your AI agent from a generic assistant into a brand-aligned, goal-oriented teammate. But writing one can be deceptively simple — vague instructions lead to vague replies, while overly complex ones can confuse the AI or break in unexpected situations.
Here’s how to think about what separates a good prompt from a bad one:
| ✅ Good prompts | 🚫 Bad prompts |
| Are goal-oriented and clearly state what the agent is expected to do. | Are vague or open-ended, with no clear task or context |
| Include specific instructions, keywords or constraints | Ask the agent to “help” without clarifying what kind of help is expected |
| Provide context about your business, tone, or customer type | Assume the AI agent already knows how your business works |
| Anticipate what might go wrong, and guide the fallback behavior | Leave the AI agent guessing when it's unsure or in unfamiliar scenarios |
Think of your prompt like giving clear instructions to a new teammate
Imagine you’re onboarding someone new to your team. You wouldn’t just say, “Help our customers.”
You’d explain who your customers are, what kind of help they usually need, and how they should communicate.
Writing prompts for your AI agent works the same way.
If you give clear, specific guidance, your agent knows how to respond — even when conversations go off-script. If your prompt is vague or overly broad, however, the AI will struggle to meet expectations.
A good prompt is:
- Grounded in your business goal: Are you trying to support users, drive sales, qualify leads, or route conversations?
- Clear about expectations: What should the agent always help with? What topics or actions are out of scope?
- Context-aware: What’s the tone, topic, or customer persona the agent is talking to?
- Flexible, not fragile: If something unexpected happens, does the prompt give enough direction for the agent to handle it gracefully?
Writing effective prompts and instructions
AgentFlow relies on natural-language prompts to guide your AI agent’s behavior across various configurations — from overall instructions to guardrails, actions, and exit conditions. Writing specific and intentional prompts ensures your agent responds accurately, sounds on-brand, and understands how to navigate different types of conversations.
Use these best practices to craft strong prompts, with examples and where they apply:
1. Be specific about your business context
Providing clear context about your company, what the agent represents, and what it should handle helps your AI agent understand its role and generate relevant replies. This leads to more accurate and helpful conversations.
| Where to apply | “Instructions” (overall behavior) |
| ✅ Example | “You are a support agent for Cat Paradise, a premium cat grooming and spa service. Help customers book appointments and answer grooming-related questions.” |
| 🚫 Avoid | “You’re a support assistant. Help people.” |
2. Define tone, style, and level of detail
Stating how your AI agent should sound and how much detail it should give ensures consistency with your brand voice and avoids overly long or overly vague answers.
| Where to apply | “Instructions”, “Send message” |
| ✅ Example | “Use a friendly and casual tone.” |
| 🚫 Avoid | “Just sound nice.” |
3. Guide how the agent should behave in different situations
Good prompts don’t just define what to say — they clarify how the agent should react to uncertainty, edge cases, or sensitive topics. This helps reduce off-topic replies and improves reliability.
| Where to apply | “Instructions”, “Guardrails”, “Send message” |
| ✅ Example | “If the customer asks something not covered in the knowledge base, politely ask for clarification. Avoid referencing external sources or assumptions. Provide as much business-specific context as possible to ensure accurate, on-brand responses.” |
| 🚫 Avoid | “If unsure, figure it out.” |
4. Use clear, structured prompts in actions
For each action, write prompts that describe what output is expected. This helps your AI agent perform consistently across conversations.
| Where to apply | “Send message”, “Calculate lead score" |
| ✅ Example | “Recommend one service plan based on what the customer needs. Highlight the key benefit in one sentence. If multiple options apply, ask a follow-up question to narrow it down.” |
| 🚫 Avoid | “Just write a response” |
5. Write precise exit condition
The “Exit condition” field tells the AI what kind of message or signal to look out for before exiting the conversation. This isn’t the message the AI will send — it’s a natural-language rule that describes when the exit should be triggered.
Writing a clear, specific condition helps the AI detect user intent accurately and prevents it from exiting too early or staying in the conversation longer than needed.
⚠️ Note:
At the moment, “Exit conditions” can only be triggered by in-conversation messages. The AI agent cannot exit based on external triggers such as label changes, contact property updates, or other system events. If you need the flow to exit based on qualification, use “Lead score” as the criteria of the exit condition instead.
| Where to apply | “Exit conditions” → “Exit condition field” |
| ✅ Example |
|
| 🚫 Avoid |
|
6. Avoid vague or open-ended instructions
Instructions like “respond nicely” or “help the customer” don’t provide enough context. Your AI agent performs best when it has a clear purpose, role, and constraints.
| Where to apply | All prompt fields |
| ✅ Example | “You are a support assistant helping customers understand product features. Keep answers short and friendly, and ask for clarification when needed.” |
| 🚫 Avoid | “You’re here to help. Just be helpful.” |
7. Avoid over-prompting and conflicting instructions
Trying to “cover everything” in a single prompt can lead to confusing or inconsistent AI behavior. If your instructions are too long, repetitive, or contradictory, your AI agent may struggle to know what to prioritize, or may fall back to vague, generic replies.
Writing intentional prompts is not about saying more, it’s about saying the right things in the right place.
| Where to apply | All prompt fields (especially “Instructions” and “Send message”) |
| ✅ Example | Write one prompt per intent. Keep it focused on a single goal — e.g. tone, fallback behavior, or task scope. Place additional instructions in the correct fields (e.g. guardrails for edge cases). |
| 🚫 Avoid |
|
✅ Tip: Read your prompt out loud. If it sounds like you’re giving mixed signals or trying to do too much at once, simplify it or split it across fields.
Example prompt patterns by industry
Different industries have different expectations for tone, behavior, and level of support. Whether you’re setting up a support-focused agent or one designed to drive sales, the prompts you write should reflect your business goals and customer needs.
In the sections below, you’ll find ready-to-use prompt patterns tailored to two common AI agent types:
- Basic support — agents that answer FAQs and handle simple service questions (e.g. business hours, refund policy, general process)
- Sales growth — agents that assist with product discovery, lead qualification, and conversion
Each example includes guidance for the Instructions and Guardrails fields to help your AI agent stay helpful, consistent, and on-brand.
In this section, we will illustrate some of the example prompts in “Instructions” and “Guardrails” by industry.
Differences between “Instructions” and “Guardrails” in AI agent configurations
- Instructions tell the AI what role it’s playing, what it should do, and how it should sound — like setting the overall character and purpose of the agent.
- Guardrails set limits: what the AI should watch out for, avoid, or escalate. Use them to handle sensitive topics, risky questions, or anything the AI shouldn’t answer directly.
Example prompts by industry – Basic support AI agent
Use these if your agent primarily helps with customer inquiries,FAQs, or issue resolution.
| Industry | Instructions | Guardrails |
| E-commerce/Retail |
You are a support assistant for Bloom & Co, a fashion brand. Help customers check return policies, and handle common product questions. Use a friendly and empathetic tone.
|
Observe for: Complaints about delays or returns.
|
| Travel and hospitality |
You are a guest services assistant for Horizon Stays. Help guests by explaining check-in policies, and answering facility-related questions. Keep your tone calm and professional.
|
Observe for: Mentions of emergencies (e.g. lost passport, illness, accident).
|
| Education |
You are a student help assistant for LearnPlus. Support users with logging in, accessing course materials, and understanding schedules. Sound clear and encouraging.
|
Observe for: Grading disputes or certification concerns.
|
| Food and beverages |
You are a virtual front-of-house assistant for The Pasta Place. Help customers view the menu, and check order status. Be warm and concise.
|
Observe for: Questions about allergens or ingredients.
|
| Professional services |
You are a client service assistant for TrustLedger. Help explain services, and access shared files. Use a formal but helpful tone.
|
Observe for: Legal or financial advice requests.
|
| Others |
You are a general support assistant helping users reach the right department or access FAQs. Stay polite and neutral.
|
Observe for: Off-topic or unsupported questions.
|
Example prompts by industry – Sales Growth AI agent
Use these if your agent primarily helps with generating leads, recommending products or services, qualifying interest, or nudging users toward conversion.
Industry |
Instructions |
Guardrails |
E-commerce/Retail |
You are a sales assistant for Bloom & Co. Help customers find products, suggest upsells, and promote discounts. Ask about preferences and suggest matching items. Keep the tone energetic and helpful.
|
Observe for: Hesitation or price-related concerns.
|
Travel and hospitality |
You are a booking assistant for Horizon Stays. Help guests discover packages, upgrades, and personalized experiences. Offer tailored recommendations based on travel dates or group size.
|
Observe for: Guest messages that express uncertainty or mention specific limitations (e.g. “not available”, “fully booked”, “I’m not sure what to choose”).
|
Education |
You are an enrollment assistant for LearnPlus. Help new learners explore relevant courses based on their goals. Recommend bundles or certifications. Sound encouraging and informed.
|
Observe for: Budget or time commitment concerns.
|
Food and beverages |
You are a promotional assistant for The Pasta Place. Suggest bestsellers, seasonal dishes, and add-ons. Encourage customers to complete their order. Use a light and upbeat tone.
|
Observe for: Customers saying they’re just browsing or unsure.
|
Professional services |
You are a lead qualifier for TrustLedger. Help potential clients understand the value of services and encourage them to schedule a consultation. Ask about business needs. Remain confident but respectful.
|
Observe for: Vague or non-committal messages (e.g. “just browsing”, “not sure”, “maybe later”).
|
Others |
You are a discovery assistant helping users explore your company’s services and match them with the right offering. Keep the tone consultative and friendly. Ask clarifying questions before recommending anything.
|
Observe for: Disinterest or rejection of suggestions.
|
Frequently asked questions
“I don’t know what to write in the prompt box. Where do I even start?”
Start by thinking about what you want your AI agent to do — not how to phrase it. Ask yourself:
- What is the goal of this flow (support, booking, sales, etc.)?
- What kind of questions will users ask?
- What should the agent always help with?
Write it as if you’re briefing a new teammate — clear, direct, and with just enough context to do the job.
“How much detail do I need to give?”
Give only what’s relevant to the task. For example, instead of saying: “Handle refunds”
- Try: “Help customers understand our refund policy for online orders. If the refund involves in-store purchases, direct them to a human agent.”
You don’t need to copy your whole refund policy — just help the agent understand where its boundaries are.
“Is this something AI can’t do, or did I just write it wrong?”
If the reply feels robotic, irrelevant, or evasive, it could be:
- A prompt that’s too vague
- A topic the AI isn’t trained to handle
- A missing fallback (For example: what to say when unsure)
Try rewriting your prompt with clearer intent. And if the topic needs human judgement (like legal disclaimers or medical advice), add a guardrail to escalate.
“How do I make the agent sound more like our team?”
Be explicit. Don’t just say: “Sound friendly.”
Instead, describe what friendly means to your brand — e.g. short replies, casual language, emojis allowed or not.
You can also use words like:
- “Use a warm, supportive tone. Avoid technical jargon. Ask follow-up questions if needed.”
The more you spell it out, the more consistent your agent will sound.
“When the agent goes off, how do I fix it?”
Start by identifying what went wrong — then make a small, targeted change. You don’t need to rewrite the whole prompt.
Here’s how to break it down:
What’s going wrong |
What to check or adjust |
The reply is off-topic or irrelevant |
Review the “Instructions”.
|
The reply is inaccurate or confusing |
Check your data source (AI Agent Knowledge sase). If the answer isn’t in there, add it. If it is but the agent handled it poorly, update your “Guardrails” to clarify how it should respond.
|
The agent freezes, guesses, or give a generic fallback |
Check if the “Guardrails” and “Exit condition” include a fallback logic (for example: what to do if the confidence score is low).
|
Are there any privacy or security practices I should follow when writing prompts?
Yes. While prompts help your AI agent understand how to behave, they should never include sensitive or private user data. Keep your prompts safe, general, and responsibility-aware.
Best practices:
- Never include personal identifiable information (PII) like names, phone numbers, account IDs, or addresses directly in any prompt field.
- Use guardrails to block or escalate sensitive topics such as passwords, payment details, medical history, or legal issues.
- Limit the scope of your prompts: avoid overloading them with unnecessary context that could expose private business logic or customer workflows.
- Test your agent for edge cases: try asking sensitive or restricted questions and confirm that the AI responds appropriately (or doesn’t respond at all).
Prompts guide behavior — but your business is responsible for what the AI can and can’t answer. Be intentional, and always protect user trust.