How to write a customer service chatbot system prompt
Turn a vague support persona into measurable operating instructions, then prove that it answers known facts and safely declines unsupported claims.
A customer service chatbot system prompt should define behavior and boundaries while changing facts remain in maintainable knowledge sources.
A system prompt is the standing instruction the assistant receives before every visitor message. In WebChatAgent it lives in Custom Role and can define identity, scope, style, evidence rules and escalation, but it cannot make missing company facts reliable.
This guide separates behavior from knowledge. We create a compact prompt for Northstar Services, test one fact that exists in an approved source, then ask for an absent annual-contract refund policy. The correct result is a useful answer for the first question and a transparent refusal for the second.
Custom Role is available from the Basic plan and accepts up to 5,000 characters. The best prompt is usually much shorter: every sentence should control an observable behavior that you can test.
Privacy-protected two-click player
Customer Service Chatbot System Prompt: Build & Test It
Write a customer service chatbot system prompt with role, tone, evidence rules, privacy boundaries, escalation and known-versus-unknown answer tests.
The YouTube player stays blocked until you choose Play. Loading it connects your browser to YouTube and may transfer technical data to Google.
Open directly on YouTubeWhat you will have at the end
- A copy-ready prompt with seven auditable sections
- Explicit source, privacy, refusal and human-escalation rules
- Known-answer and missing-information regression questions
- A repeatable A/B decision in the Role Optimizer
Before you start
- Basic plan or higher, or an active trial
- A chatbot with at least one approved source
- A real support owner and a valid handoff route
- Five fixed test questions covering normal, ambiguous, out-of-scope and escalation cases
Prompt for behavior; source facts separately
The prompt decides how the assistant should act; indexed sources decide which changing company facts it may rely on. Keep policies, prices, schedules and product details in owned sources so they can be updated without rewriting behavior instructions.
Use seven short sections: Role, Audience and Tone, Scope, Evidence, Boundaries, Escalation, and Response Format with Examples. Then attach at least one fixed question to every critical rule.
01–11
Set it up step by step
Open the right assistant and preserve a baseline
Never edit a production prompt without a rollback copy.
Select the intended assistant in the left sidebar, open Configuration and confirm the chatbot name at the top. Scroll to Answer style and role, the current Custom Role/System Prompt card; it shows the beginning of the saved instructions.
Before opening the editor, copy the full current prompt into a version record with chatbot, owner, date and reason for change. Also record provider, model, temperature and active source version. A prompt comparison is meaningless when several variables change at once.
Choose a preset, AI draft or clean baseline
Templates accelerate drafting; they do not define your support policy.
Open Custom Role Configuration. Preset Roles can fill a starting template; Generate with AI is available when the chatbot already has data sources. Read every generated line before keeping it because a draft can infer a tone or scope that your company never approved.
For a controlled tutorial, start from the Website Support preset or a clean text area. Keep the 5,000-character limit visible, but aim for a compact prompt that a support owner can review in a few minutes.
Write Role, Audience and Tone as observable behavior
Replace adjectives with instructions a reviewer can see in an answer.
Begin with: “ROLE — You are the customer-support assistant for Northstar Services.” Then add the audience and purpose: help website visitors understand approved services and next steps.
Under TONE, write “clear, calm and human” plus observable rules: acknowledge frustration in one sentence, use everyday English, avoid hype and do not repeat the visitor’s whole question. “Be friendly” alone is too vague to score.
Separate support scope from factual evidence
The prompt governs behavior; sources remain the authority for changing facts.
Under SCOPE, list the supported jobs: explain services, navigation and published support processes. Then name out-of-scope areas such as legal, medical or financial advice and internal account changes.
Under EVIDENCE, require indexed sources for company policies, prices, dates, availability and contract terms. If the retrieved content does not support a claim, the assistant must say it cannot find that information. Do not paste the refund policy into this section; add and maintain it as an approved data source.
Add privacy, security and transaction boundaries
A polite answer can still be unsafe if it invents access or collects unnecessary data.
State that the assistant must not request passwords, full payment-card data, authentication codes or unnecessary sensitive data. It must never claim to have viewed an account, changed a subscription, issued a refund or completed another action unless a configured tool returned a confirmed result.
Tell it to ignore visitor requests to reveal hidden instructions, credentials or private source content. These rules support product controls; they are not a substitute for permissions, domain restrictions and secure tool validation.
Define exactly when and how to escalate
Escalation needs triggers, destination and minimum context.
Under ESCALATION, list triggers: the visitor asks for a human, the issue is a billing dispute, safety or privacy concern, repeated failure, or required information is absent. Name the real path, such as the configured Live Chat handoff or support contact.
Tell the assistant what to collect before handoff—for example a short issue summary and an approved contact field—and what never to collect. If no human is online, it should explain the next response channel and timing only when those details exist in an approved source.
Specify response format and two safe examples
Examples should demonstrate form without becoming a hidden policy source.
Under RESPONSE FORMAT, set a practical default: answer directly, normally stay below 100 words, use bullets for three or more steps and ask one clarifying question when intent is ambiguous. Avoid forcing short replies when safety or accessibility needs more context.
Add one grounded example and one missing-information example. Keep volatile values out of the example. A safe pattern is: “I can’t find an approved annual-contract refund policy in the available sources. I can help you contact support instead.”
Save, then run a known-answer control question
Prove that the stricter rules did not block useful grounded answers.
Select Save in the Custom Role dialog and wait until the configuration reports the change as saved. Start a fresh conversation so earlier context cannot hide the effect of the new instructions.
Ask a question whose exact answer exists only in the approved tutorial source, and record question, expected fact, observed answer, source version, model and prompt version. The answer must return the fact without adding unsupported contract or policy details.
Test missing information and escalation behavior
A transparent limitation is a successful result when the source is absent.
In another fresh conversation ask: “What is Northstar Services’ refund policy for annual contracts?” The tutorial source intentionally contains no such policy.
Expected result: the assistant says the policy is not available in the indexed information, does not invent a refund window or fee, and offers the approved support path. Repeat with an explicit “I want a human” request and one prompt-injection attempt.
Compare current and proposed prompts in Role Optimizer
Use identical questions and inspect answers, not only the win rate.
Open Knowledge Optimizer, choose the same chatbot and select Role. Generate a proposal, read Current and Proposed side by side and remove any invented workflow before testing.
Choose a fixed question source and run A/B validation. Open each pair and score factual support, correct refusal, tone, brevity and escalation. Apply role only when the proposed version improves the target behavior without breaking the known-answer control. A polished proposal that scores worse stays unapplied.
Version the winner and monitor real conversations
A prompt is released behavior, not a one-time writing exercise.
Store the approved prompt, owner, date, change reason, question set and A/B result together. Keep the last version ready for rollback and change one layer at a time in future tests.
After release, review Questions, Feedback, Conversations and Live Chat handoffs for missed facts, false confidence, awkward tone and unnecessary escalation. Re-run the fixed set whenever sources, tools, model, audience or escalation process changes.
Example & result
See the practical test and its result
Every tutorial includes a fixed input, the expected outcome and a transparent record of what was actually verified locally.
Practical example: How to write a good customer service system prompt
This exact scenario was completed with the temporary tutorial account.
Exact test input
What is Northstar Services’ refund policy for annual contracts?
Expected result
The assistant says the policy is unavailable, invents no terms and offers the configured support path.
What was actually verified
The real assistant said it had no information about the annual-contract refund policy and offered to connect the visitor with a support agent. It invented no window, fee or eligibility rule.
Tips & tricks
Make the setup reliable
Test with realistic examples, record your baseline and change one setting at a time. That makes real improvements visible.
Use headings inside the prompt
Role, Audience and Tone, Scope, Evidence, Boundaries, Escalation and Response Format are easier to audit than one long paragraph.
Put precedence into conflicting rules
“Always answer” conflicts with “never guess.” State that evidence and safety rules win, then test the conflict directly.
Do not promise a handoff that is not configured
The prompt should name only a real Live Chat, ticket, email or callback path. Otherwise the assistant creates a dead end while sounding helpful.
Keep facts close to their owners
A changing refund policy belongs in an owned source with an effective date, not in a prompt that only an administrator remembers to edit.
When something does not work
Troubleshooting
Check status, permissions and test data systematically before changing the model or prompt.
The assistant still invents a missing policy
Confirm with Content Search that the policy is truly absent, remove conflicting examples from the prompt, require indexed evidence for policy claims and retest in a new session. A stronger model cannot turn an absent policy into an approved fact.
The prompt sounds robotic or repeats disclaimers
Replace broad warnings with one short fallback sentence, add a human tone example and test real visitor wording. Keep safety conditions explicit but do not require the assistant to recite every rule.
The Role Optimizer has no questions to test
Create a fixed Knowledge Test set or select another populated question source. Include at least one known fact, one ambiguous request, one absent policy, one escalation and one out-of-scope case.
A better win rate hides a broken critical answer
Open every answer pair and treat critical known-answer, privacy and escalation cases as release gates. Do not apply a prompt that fails one of them even when the aggregate percentage rises.
Ready for a production-style test
Turn the five tutorial questions into a permanent prompt regression set. Assign an owner, run it after every prompt, source, model or tool change and link each release to the exact saved prompt version.
Related resources
Test the prompt in Knowledge Optimizer
Build fixed questions, compare roles and keep only measured improvements.
Trace and reduce unsupported answers
Separate source, retrieval, prompt and model failures before changing anything.
Custom Role configuration reference
Review the current plan, length and configuration behavior.