Good chatbot prompts do more than ask an AI model to “be helpful.” They define the assistant’s role, information boundaries, response format, escalation rules, and success criteria. This guide provides reusable AI chatbot prompt templates for customer support, internal search, FAQ responses, lead qualification, summarization, and safe refusal handling, along with practical methods for adapting and maintaining them.
Overview
An AI chatbot prompt is the operating brief that guides how an assistant responds. In a simple FAQ bot, the prompt may establish a tone and instruct the model to answer from approved content. In a more advanced knowledge base chatbot or RAG chatbot, it can also explain how to use retrieved passages, handle conflicting information, cite sources, and transfer difficult cases to a person.
The most reliable prompts separate stable instructions from changing business content. Stable instructions include the assistant’s role, scope, safety rules, and output format. Changing content includes product details, policies, prices, availability, and internal procedures. Those details should generally come from a maintained knowledge base, document store, or retrieval workflow rather than being permanently written into the prompt.
A reusable prompt should answer five questions:
- Who is the assistant helping?
- What sources may it use?
- What should it do when the answer is incomplete?
- What should the response look like?
- When should it ask a question or involve a human?
These principles apply whether you are creating an AI support chatbot, an internal AI assistant for teams, an AI chatbot for a website, or a custom chatbot connected through a chatbot API. For implementation details such as authentication, webhooks, and streaming, see the AI Chatbot API Integration Guide.
Template structure
Use the following structure as a starting point. It is intentionally modular so that developers and administrators can adapt individual sections without rewriting the entire prompt.
ROLE
You are [assistant name], a [type of assistant] for [audience]. Your job is to [primary outcome].
SCOPE
You can help with [approved topics]. You cannot provide authoritative answers about [excluded topics].
SOURCE RULES
Use the supplied knowledge base or retrieved context as the primary source.
If the source does not contain enough information, say so clearly.
Do not invent facts, policies, links, deadlines, or account-specific details.
RESPONSE METHOD
1. Identify the user’s goal.
2. Use relevant approved context.
3. Answer directly in [preferred format].
4. State important assumptions or missing information.
5. Ask one focused follow-up question when necessary.
ESCALATION
Recommend human assistance when [defined conditions].
Before handing off, summarize the issue and collect [approved details].
STYLE
Be [tone]. Use [format preferences]. Avoid [undesired behaviors].The role section should be narrow enough to guide decisions. “You are a helpful assistant” is too broad for most business workflows. “You help customers understand documented account setup and troubleshooting steps” gives the model a clearer boundary.
The source rules are particularly important for a document chatbot or knowledge base chatbot. A useful instruction is to distinguish between retrieved evidence and general reasoning: the assistant may explain information in the source, but it should not present an unsupported assumption as a confirmed policy. If citations are part of the product experience, specify the citation format and require the assistant to attach sources only when they support the answer.
Output rules should match the workflow. A customer-facing answer may need short paragraphs and numbered steps. An internal search assistant may need a direct answer followed by source titles. A lead qualification bot may need structured fields that your CRM or chatbot API can process consistently.
How to customize
Start by defining the decision the chatbot must make, not by choosing a personality. For support, the decision may be whether to answer, troubleshoot, collect information, or escalate. For internal search, it may be whether the retrieved documents actually answer the question. For lead qualification, it may be whether the visitor has a relevant use case and a suitable next step.
Customer support triage
Tell the AI support chatbot to classify the request before responding. Useful categories might include how-to question, known issue, account-specific request, billing concern, and urgent escalation. Instruct it to provide documented steps for routine issues, avoid requesting sensitive information in chat, and explain what information a human agent needs when escalation is appropriate.
Internal knowledge retrieval
For an internal AI assistant, instruct the model to prioritize retrieved company documents and identify the relevant document or section. Add a rule for conflicting documents: acknowledge the conflict instead of silently selecting one. You can also require the assistant to distinguish between a confirmed answer, a likely interpretation, and an unanswered question.
FAQ and website answers
A website chatbot should answer common questions in a compact format. Set a practical length limit, include a link when an approved destination exists, and ask only one follow-up question at a time. If you plan to embed the chatbot on a website, test the prompt with short, vague, misspelled, and multi-part questions rather than only polished examples.
Lead qualification
For lead qualification, define the minimum information to collect, such as use case, team size, current workflow, desired integration, or implementation timeframe. Explain why each question is asked and avoid collecting unnecessary personal information. The prompt should instruct the bot to summarize the conversation accurately and offer a relevant next step without making promises about pricing, availability, or capabilities that are not confirmed.
Keep prompt variables explicit. For example, use placeholders such as {{company_name}}, {{retrieved_context}}, {{support_channel}}, and {{handoff_url}}. This makes the prompt easier to version, test, and reuse across environments.
Examples
AI support chatbot prompt
You are a support assistant for [product]. Help users resolve documented product questions.
Use the supplied help-center content as your primary source. Give numbered steps when troubleshooting.
If the content does not answer the question, say that you cannot confirm the answer and offer human support.
Do not invent account status, refunds, security findings, or policy exceptions.
If the user reports a possible security issue, stop troubleshooting and direct them to [approved channel].
End with one question only when more information is required.Internal search prompt
You are an internal knowledge assistant. Answer employees using the retrieved company documents.
Begin with the direct answer, then list the relevant source title or section.
If the sources disagree or do not answer the question, explain the limitation and identify what should be verified.
Do not treat a draft, example, or outdated document as current policy unless the context explicitly confirms it.Lead qualification prompt
You qualify visitors for a product consultation without pressuring them.
Ask one question at a time about the visitor’s use case, team, current process, and desired integration.
Summarize their answers in these fields: use_case, team_context, current_workflow, integration_need, next_step.
Do not claim that a feature, price, timeline, or integration is available unless approved information confirms it.
When the visitor is ready, offer the approved contact or booking path.Safe refusal and uncertainty handling
If the request is outside your scope, explain the limitation briefly and suggest the closest supported option.
If the answer depends on information you cannot access, say what is missing.
Never fill gaps with invented details.
If a human review is needed, explain the reason and provide the approved handoff path.These examples can be combined, but avoid creating one oversized instruction block for every use case. A support bot and a lead qualification bot have different goals, failure modes, and output requirements. Separate prompts or clearly separated modes are usually easier to evaluate.
When to update
Revisit chatbot prompts whenever the workflow, source material, or failure pattern changes. A prompt that worked during an initial pilot may need adjustment after new documents are added, a support process changes, or users begin asking questions that were not represented in testing.
Review the prompt after:
- Adding a new product area, knowledge source, integration, or escalation channel.
- Changing support ownership, approval steps, or the desired response format.
- Observing hallucinations, unsupported certainty, repetitive questions, or incorrect handoffs.
- Changing the retrieval pipeline, document chunking, metadata, or citation behavior.
- Introducing a new model, model setting, chatbot API, or website interface.
- Receiving repeated user feedback about unclear or incomplete answers.
Maintain a small test set of real or carefully anonymized questions. Include straightforward requests, ambiguous wording, missing information, conflicting documents, adversarial instructions, and escalation cases. Compare prompt versions against the same test set so improvements are based on observed behavior rather than impressions.
For a practical workflow, store each prompt with an owner, version number, purpose, approved variables, and last review date. Pair prompt changes with knowledge-base maintenance: a better instruction cannot compensate for missing or contradictory source content. If your project uses retrieval, the prompt engineering guide for knowledge bots and RAG chatbot implementation guide can help connect prompt design with the wider answer pipeline.
To put this library into practice, choose one workflow, copy the closest template, replace every placeholder, and test it with at least ten representative questions. Record where the assistant is uncertain, overconfident, too long, or too eager to escalate. Then revise the smallest section that addresses the problem. This incremental approach keeps AI chatbot templates maintainable while giving your team a clear path from an initial prompt to a dependable production workflow.
