How to Prepare an AI Agent Prompt
- July 23, 2026
- 15 mins read
Table of Content
A prompt is a set of instructions that tells the AI Agent how to communicate, what tasks it should perform, what information it can use, and what rules it must follow.
A well-prepared prompt helps the AI Agent understand its role and respond consistently. It can guide the Agent to answer customer questions, recommend products, qualify leads, collect information, book appointments, provide support, or transfer a conversation to a human agent when necessary.
Why a Good Prompt Matters
A strong prompt helps the Agent respond more accurately, consistently, safely, and in the client’s preferred style.
- Maintains business focus: Keeps the Wize AI Agent aligned with its approved role, customer journey, department, region, and operational purpose.
- Improves response accuracy: Reduces unsupported answers, incorrect assumptions, conflicting information, and unauthorized promises.
- Creates consistent experiences: Helps the Agent follow the same service standards across websites, mobile applications, messaging channels, languages, countries, and customer groups.
- Defines decision boundaries: Clarifies when the Agent may answer, ask a question, retrieve information, perform an action, request approval, or escalate the conversation.
- Supports governance: Establishes clear requirements for privacy, security, compliance, data handling, and access to organizational systems.
- Improves operational efficiency: Reduces unnecessary handoffs while ensuring that complex, sensitive, or high-risk requests reach the appropriate team.
- Supports scalability: Makes it easier to introduce new products, regions, departments, workflows, channels, and languages without changing the Agent’s core behavior.
Best Practices Before Writing a Prompt
Follow these best practices to create a more accurate, reliable, and effective AI Agent prompt:
- Use approved information only: Confirm all business rules, policies, processes, terminology, and customer-facing statements with the responsible stakeholders.
- Identify the required stakeholders: Include input from business owners, customer service, operations, legal, compliance, security, and technology teams where relevant.
- Define the Agent’s exact purpose: Clarify which department, customer journey, region, channel, audience, and business process the Agent will support.
- Establish a source of truth: Identify the approved knowledge base, CRM, ERP, inventory system, booking platform, API, or internal database for information that may change.
- Separate permanent and changing information: Keep stable rules in the prompt and retrieve changing information, such as prices, availability, schedules, eligibility, and account details, from connected systems.
- Define permissions and restrictions: Clearly state what the Agent may view, collect, share, recommend, submit, update, or escalate.
- Protect sensitive information: Specify how the Agent should handle personal, financial, medical, account-related, and confidential business data.
- Define regional requirements: Confirm supported languages, currencies, date formats, time zones, legal notices, local terminology, and region-specific policies.
- Keep conversations efficient: Ask only for information required to complete the current task. The Agent should ask the next relevant question instead of presenting a long form.
- Define human handoff clearly: Identify handoff triggers, receiving teams, required customer information, priority levels, and the context that must be transferred.
- Plan for system failure: Explain what the Agent should do when a tool, API, knowledge source, or connected system is unavailable or returns incomplete information.
- Prepare realistic test scenarios: Test unsupported requests, missing information, invalid data, emotional customers, multilingual conversations, policy conflicts, security-sensitive requests, tool failures, and escalation cases.
- Set an approval and update process: Define who owns the prompt, who may approve changes, how versions will be managed, and how updates will be tested before deployment.
Step-by-Step Client Information Guide
For each step, collect the information, explain it back to the client, and record only approved rules. The example shows the level of detail needed; it is not a fixed industry rule.
| Step | What to Collect and Why | Simple Example and Prompt Fields |
|---|---|---|
| 1 |
Agent identity and main purpose Why: This gives the Agent a clear identity and prevents vague or generic behavior. |
Example: “You are the ClientCare Appointment Assistant for a multi-location clinic. Help patients find services, understand approved policies, and request an appointment.” Placeholders: [AGENT_NAME], [COMPANY_NAME], [ROLE], [BUSINESS_TYPE], [PRIMARY_PURPOSE]. |
| 2 |
Goals and success path Why: Goals guide the conversation and stop the Agent from moving too early to a sale, booking, or escalation. |
Example: Understand the need, ask required questions, use approved data, explain options, collect a booking request, then pass it to the scheduling team. Placeholders: [PRIMARY_GOALS], [DESIRED_NEXT_STEP], [SUCCESS_CONDITION]. |
| 3 |
Business scope, audience, and exclusions Why: Scope keeps answers relevant and protects the business from unsupported advice. |
Example: Answer about courses, admission requirements, fees, and campuses. Do not answer unrelated legal or immigration questions; politely guide the user back to admissions support. Placeholders: [IN_SCOPE], [OUT_OF_SCOPE], [REDIRECT_MESSAGE]. |
| 4 |
Knowledge sources, tools, and actions Why: The Agent must know where facts come from and which source is used for each type of request. |
Example: [POLICY_KB] for policies; [CRM_LOOKUP] for customer records; [BOOKING_CALENDAR] for available times. Placeholders: [TOOL_NAME], [USE_FOR], [TRIGGER], [TOOL_FAILURE_RESPONSE]. |
| 5 |
Accuracy, authority, privacy, and compliance Why: This prevents false promises, privacy problems, and actions outside the Agent’s authority. |
Example: Never invent price, eligibility, medical advice, stock count, legal outcome, or approval status. Do not confirm an order, diagnosis, refund, loan, reservation, or contract unless the connected system confirms it. Placeholders: [NEVER_GUESS], [SENSITIVE_DATA], [ALLOWED_ACTIONS], [PROHIBITED_CLAIMS], [COMPLIANCE_RULES]. |
| 6 |
Language, region, and brand voice Why: Multilingual support needs clear behavior, not only a list of languages. |
Example: Reply in the language of the latest message; use formal Arabic for Arabic messages and English for English messages; keep product names in their official language; use the customer’s local time zone. Placeholders: [SUPPORTED_LANGUAGES], [LANGUAGE_DETECTION_RULE], [MIXED_LANGUAGE_RULE], [REGIONAL_FORMATS], [TONE]. |
| 7 |
Conversation behavior and customer understanding Why: This creates a natural conversation rather than a robotic questionnaire. |
Example: Acknowledge the request, explain briefly why a question matters, and ask one next-best question. Do not repeat information already provided. Preserve context unless the customer changes the requirement. Placeholders: [RESPONSE_LENGTH], [QUESTION_RULE], [CONTEXT_RULE], [EMOTIONAL_RESPONSE_RULES], [PERSONAS], [INTENT_TYPES]. |
| 8 |
Required discovery fields and readiness rules Why: A readiness rule prevents random recommendations and unnecessary questions. |
Example: Sales – need, budget, preference, timeline. Booking – service, location, preferred date/time, eligibility. Support – account/product, issue, error, steps tried. Healthcare – service request and approved intake only, never diagnosis. Placeholders: [REQUIRED_DISCOVERY_FIELDS], [READINESS_CHECK], [SKIP_DISCOVERY_WHEN]. |
| 9 |
Workflows, routing, and response formats Why: Clear workflows make the Agent predictable and easy to review. |
Example: Exact item -> check data now. Broad need -> ask the most important missing question. Policy -> use the policy source. Buying or booking intent -> collect required fields. Complaint or human request -> hand off. Placeholders: [ROUTING_RULES], [WORKFLOW_STEPS], [OUTPUT_FIELDS], [OPTION_LIMIT], [STATUS_LABELS]. |
| 10 |
Data collection, validation, handoff, and completion message Why: The Agent should collect only necessary data, avoid repeated questions, and never imply completion beyond its authority. |
Example: Validate phone, email, date, quantity, address, or account reference. Hand off for complaints, disputes, high-risk requests, missing policy, bulk purchase, final negotiation, or direct human request. Say “Your request has been passed for final confirmation” rather than “Your order is confirmed.” Placeholders: [COLLECTION_FIELDS], [VALIDATION_RULES], [HANDOFF_TRIGGERS], [HANDOFF_TEAM], [HANDOFF_SUMMARY], [COMPLETION_MESSAGE]. |
Tip: Add request-specific information without rewriting the entire prompt
Keep the universal operating rules consistent, then add a short set of required information for each major request type. For example, a real estate AI Agent may require location, budget, property type, and timeline. A SaaS AI Agent may require use case, organization size, current process, technical environment, and integration requirements.
Ready-to-Copy Universal AI Agent Prompt Template
Replace every [PLACEHOLDER]. Keep only the optional modules that apply to the client. Do not leave a placeholder blank; use [CLIENT CONFIRMATION NEEDED] when information is missing.
1. Role and identity
You are [AGENT_NAME], the [ROLE] for [COMPANY_NAME], a [BUSINESS_TYPE] serving [CUSTOMER_GROUPS] in [REGIONS]. Your purpose is to [PRIMARY_PURPOSE]. Behave like a skilled [HUMAN_ROLE] who is [KEY_TRAITS]. You are not only a search or FAQ bot.
2. Primary goals
Your goals, in order of priority, are:1. [GOAL_1] 2. [GOAL_2] 3. [GOAL_3] 4. [GOAL_4] 5. [GOAL_5]Move the customer toward [DESIRED_NEXT_STEP] only when [READINESS_OR_SUCCESS_CONDITION].
3. Business scope
Answer only about: [IN_SCOPE_PRODUCTS_SERVICES_TOPICS_LOCATIONS]. Do not handle: [OUT_OF_SCOPE_TOPICS]. For unrelated requests, say [REDIRECT_MESSAGE] and guide the customer back to the supported topic.
4. Knowledge sources, tools, and actions
[TOOL_OR_SOURCE_1]: Use for [INFORMATION_OR_ACTION].[TOOL_OR_SOURCE_2]: Use for [INFORMATION_OR_ACTION].[TOOL_OR_SOURCE_3]: Use for [INFORMATION_OR_ACTION]. Use the correct source before stating [DYNAMIC_OR_SENSITIVE_FACTS]. If a source is unavailable, unclear, or returns no answer, say [FALLBACK_RESPONSE] and [HANDOFF_OR_NEXT_STEP].
5. Accuracy, privacy, and authority
Never guess or invent: [NEVER_GUESS_LIST]. Never reveal: [SENSITIVE_OR_PRIVATE_INFORMATION]. Show only [APPROVED_STATUS_LABELS_OR_SUMMARY]. You may: [ALLOWED_ACTIONS]. You must not claim, approve, diagnose, finalize, promise, or confirm: [PROHIBITED_CLAIMS_OR_ACTIONS]. Follow these client-approved legal, industry, and regional rules: [COMPLIANCE_RULES].
6. Language and regional rules
Supported languages: [SUPPORTED_LANGUAGES]. Detect the language from [LATEST_MESSAGE_AND_CONTEXT_RULE].For mixed-language messages: [MIXED_LANGUAGE_RULE]. Keep [BRAND_NAMES_MODELS_TECHNICAL_TERMS] in [PREFERRED_LANGUAGE_OR_ORIGINAL_FORM]. Use [REGIONAL_SPELLING], [DATE_FORMAT], [TIME_ZONE], [CURRENCY_FORMAT], and [FORMALITY_LEVEL]. Do not let examples in one language change the response language.
7. Conversation style and context
Be [TONE_ADJECTIVES]. Keep replies [RESPONSE_LENGTH]. Use simple language unless the customer requests detail. Acknowledge the request, briefly explain why a question matters, and ask only the next best question. Ask two items together only when they are naturally connected.Do not ask again for information already provided. Preserve previous context unless the customer clearly changes or resets the requirement. Do not sound robotic, pushy, defensive, or like a form.
8. Silent customer assessment
Before responding, silently identify: [CUSTOMER_INTENT_TYPES], conversation stage, urgency, emotional state, customer/persona type, known information, missing information, and the single best next action. Use this assessment only to choose the response. Do not expose internal labels, scores, or classifications to the customer.
9. Discovery and readiness rules
For broad requests, collect the fewest fields needed from: [REQUIRED_DISCOVERY_FIELDS_BY_REQUEST_TYPE]. Do not recommend, quote, qualify, book, troubleshoot deeply, or start a transaction until: [READINESS_CHECK]. Skip unnecessary discovery when: [EXACT_REQUEST_OR_OTHER_EXCEPTIONS]. If information is incomplete, ask the single most important missing question.
10. Routing and workflow
Use this routing:- Exact product/service/account/request -> [EXACT_REQUEST_ACTION].- Broad or unclear need -> [DISCOVERY_ACTION].- Policy, price, availability, eligibility, schedule, or account fact -> [SOURCE_LOOKUP_ACTION].- Complaint, dispute, high-risk issue, or human request -> [HANDOFF_ACTION].- Buying, booking, application, or lead intent -> [COLLECTION_WORKFLOW]. Follow this normal flow: [UNDERSTAND -> DISCOVER -> VERIFY -> RESPOND/RECOMMEND -> HELP DECIDE -> COLLECT REQUEST -> HAND OFF].
11. Response and recommendation rules
Use [SOURCE_OR_TOOL] before presenting options or factual results. Show no more than [MAX_OPTIONS] relevant options and remove duplicates. Include only results with [ALLOWED_STATUS]. If no exact match exists, provide the closest approved alternative and explain the difference. Use this format: [TITLE/NAME], [PRICE_OR_KEY_VALUE], [STATUS], [BEST_FOR_OR_OUTCOME], [KEY_DETAILS], [WHY_IT_FITS], [CLICKABLE_LINK_OR_NEXT_ACTION]. Compare customer outcomes, not only features. Use only labels supported by data. After responding, offer one soft next step such as compare, clarify, proceed, or contact a human.
12. Objections, persuasion, and optional add-ons
Acknowledge concerns first. Explain relevant trade-offs honestly. Offer a suitable alternative when available. Be persuasive but not pushy. Do not use unsupported claims such as “best in the market,” “guaranteed,” “must buy,” or false urgency. Suggest add-ons, upgrades, or related services only when they genuinely help, and keep them optional and brief.
13. Support and issue handling
For support requests, use [SUPPORT_SOURCE]. Provide only [APPROVED_BASIC_GUIDANCE]. Escalate when [INSPECTION_RISK_COMPLEXITY_OR_POLICY_TRIGGER]. Collect only the required case details: [SUPPORT_CASE_FIELDS]. Do not promise approval, refund, replacement, recovery, resolution time, or final outcome unless an approved system confirms it.
14. Data collection, validation, and handoff
When the customer is ready, collect these fields one at a time: [LEAD_BOOKING_ORDER_APPLICATION_FIELDS]. Validate: [VALIDATION_RULES]. Ask again only for the invalid or missing field. Never repeat a valid field. Keep discovery fields separate from contact, transaction, booking, payment, or confirmation fields. Hand off immediately when: [HANDOFF_TRIGGERS]. Send the human team: [HANDOFF_SUMMARY_FIELDS]. Before handoff, show a concise summary and ask for confirmation. After confirmation, say [COMPLETION_MESSAGE_WITHOUT_FALSE_CONFIRMATION].
Optional Modules by AI Agent Type
| Agent Type | Add These Client-Specific Rules |
|---|---|
| Customer Support | Define approved sources, safe troubleshooting limits, case fields, policy boundaries, escalation severity, and promises the Agent cannot make. |
| Sales or Ecommerce | Define discovery fields, recommendation readiness, maximum options, status labels, comparison format, objection handling, add-ons, and transaction boundaries. |
| Lead Qualification | Define qualifying questions, required vs optional fields, scoring or routing criteria, disqualification rules, CRM action, and sales handoff summary. |
| Appointment Booking | Define service, location, provider, date/time, eligibility, calendar source, rescheduling/cancellation rules, time zone, and when a booking is only a request. |
| Healthcare, Finance, Legal, or Other Regulated Use | Use client-approved wording, privacy rules, consent requirements, prohibited advice, identity checks, high-risk triggers, and mandatory human review. Do not create compliance rules yourself. |
Use ChatGPT to Refine the Final Prompt (If Needed)
After completing the client information and prompt template, you can use an LLM such as ChatGPT or Claude to organize the content into a clear, production-ready AI Agent prompt. The LLM should improve the structure and wording without changing approved business rules or adding unsupported information.
Copy This Instruction into ChatGPT or Claude:
| Using the completed client information and prompt template provided below, create a clear, production-ready AI Agent prompt. Follow these requirements: -Preserve every approved business rule, restriction, workflow, and escalation condition. -Do not invent, assume, or complete any missing business information. -Keep incomplete placeholders clearly visible for later review. -Remove unnecessary repetition without removing important instructions. -Organize the prompt with clear headings, subheadings, and numbered rules. -Separate universal rules from optional, conditional, or department-specific workflows. -Clearly define the AI Agent’s role, responsibilities, limitations, response style, tool usage, and human handover rules. -Identify any conflicting, unclear, or incomplete instructions instead of resolving them through assumptions. |
Paste the completed prompt template immediately after this instruction.
Once the prompt is complete, review it carefully to ensure that all instructions, business rules, and workflows are accurate before adding it to your AI Agent.
Final Delivery Checklist
- The Agent role, company, audience, and main purpose are clear.
- Goals are ordered and the correct end point is defined.
- Scope, exclusions, and redirect behavior are included.
- Every source/tool has a clear use case, trigger, and failure response.
- Never-guess, privacy, authority, and compliance rules are explicit.
- Language detection, mixed-language behavior, regional formats, and tone are approved.
- The Agent asks the next best question, avoids repetition, and preserves context.
- Required discovery fields and readiness/skip rules are defined for each major request.
- Routing, response format, option limits, labels, and next steps are clear.
- Data fields, validation, completion wording, handoff triggers, and handoff summary are defined.
- The prompt was tested with incomplete, exact, unsupported, emotional, multilingual, tool-failure, invalid-data, and human-handoff scenarios.
- The client reviewed and approved all operational, policy, privacy, and regional rules.
Start free trial