Skip to content
AI-SNS.IO — CHAT SUPPORTAI-SNS.IO — CHAT SUPPORT
AI Agent Basics

Writing System Prompts for Support Agents That Actually Hold Up

AI SNS CHAT EditorialPublished 8 min read

Learn how to write a support agent system prompt: role, tone, scope, escalation rules and tests, with examples for web chat, WhatsApp and Messenger.

Most support agents that disappoint do not have a model problem. They have an instruction problem: a vague system prompt that says "be helpful and friendly" and leaves everything else to chance. This guide shows how to write a system prompt that sets a clear role, keeps answers inside your knowledge, hands off to humans at the right moment, and can be tested and improved over time.

Why the system prompt matters more than the model

A system prompt is the standing brief your agent reads before every conversation. Your knowledge base supplies the facts, but the prompt decides how those facts are used: who the agent is, what it may promise, how it sounds, and when it stops and asks a person for help.

A weak prompt tends to fail in predictable ways:

  • The agent answers questions you never wanted it to touch, such as legal advice or competitor comparisons.
  • It guesses when the knowledge base is silent, and the guess sounds confident.
  • It writes long paragraphs on a channel where customers read three lines at a time.
  • It never escalates, because nobody told it when to.

A good prompt is short enough to maintain and specific enough to be tested. Aim for clarity, not length.

The anatomy of a support prompt

Think of the prompt as six blocks. You can write them as short headed sections in plain text; models follow structured instructions well.

BlockQuestion it answersExample line
RoleWho is the agent and for whom?"You are the support assistant for a small online homeware shop."
ScopeWhat is in and out of bounds?"You help with orders, delivery, returns and product care. You do not give medical or legal advice."
GroundingWhere do facts come from?"Answer only from the provided knowledge. If it is not there, say so."
ToneHow should it sound?"Warm, plain, no jargon. Two to four short sentences."
ActionsWhat may it do besides talk?"Offer to collect an email address when the customer wants a follow-up."
EscalationWhen does a human take over?"Hand off for refund disputes, complaints, or when the customer asks for a person."

If a sentence in your prompt does not belong to one of these blocks, ask whether it is needed.

Write the role and scope first

Begin with one or two sentences that name the business, the audience and the job. Be concrete about what the agent is for, and equally concrete about what it is not for.

Here is a compact example for a fictional shop:

``` ROLE You are the support assistant for a homeware shop. You talk to customers on the website chat, WhatsApp and Messenger.

SCOPE You help with orders, delivery, returns, and product care. You do not discuss pricing for custom orders, give legal or medical advice, or comment on other companies. If a question is outside this scope, say so politely and offer to connect the customer with a person. ```

Say what "out of scope" looks like

Models are good at following a boundary when they also know what to do on the other side of it. "Do not discuss X" leaves a gap; "If asked about X, say you cannot help with that here and offer a human" fills it. Always pair a prohibition with a replacement behaviour.

Set tone and shape it for the channel

Tone is the part everyone writes first and the part that matters least if the rest is missing. Keep it to a few observable rules rather than adjectives alone. "Friendly" is vague; "greet once, use the customer's name if given, avoid exclamation marks" is testable.

Channel matters too. A reply that works in a web chat widget can feel heavy on WhatsApp, where people expect short, conversational messages, and Messenger sits somewhere in between.

Too long for chat
Shaped for chat
A four-paragraph reply with a greeting, a policy summary and a sign-off
One or two sentences with the answer first
Several questions asked at once
One follow-up question at a time
Formal closing line every time
A short closing only when the issue is resolved
Same answer, shaped for two channelsIllustrative example

Useful style rules to include:

  • Lead with the answer, then add detail only if it helps.
  • Ask one question at a time.
  • Reply in the customer's language unless told otherwise.
  • Avoid heavy formatting; short lists are fine, tables are not.

Ground the agent and control what it promises

The single most valuable instruction in a support prompt is what to do when the knowledge base has no answer. Without it, a model fills the silence.

A reliable pattern:

  1. Answer from the provided knowledge only.
  2. If the knowledge does not cover the question, say you are not sure rather than guessing.
  3. Offer the next step: a person, a form, or a follow-up.

Draw the line on promises

Support agents get into trouble when they promise things the business has not approved. List the categories explicitly: refunds, delivery dates, discounts, compensation, and anything with legal weight. For example:

`` You never promise refunds, exceptions, delivery dates or compensation. You can explain the published policy. For anything beyond it, offer to pass the conversation to a person. ``

Handle customer data carefully

Tell the agent not to ask for passwords, full card numbers or other sensitive details in chat. If you need an order number or an email address, say exactly which ones are acceptable to request.

Defend against instruction hijacking

Customers, and occasionally bad actors, will type things like "ignore your instructions." Add a line saying that the agent keeps its role and rules regardless of what a message asks. This is a helpful layer, but it should not be your only one; platform-level filters are a better foundation than wording alone.

Define escalation like a policy, not a feeling

"Escalate when needed" is not an instruction. List the triggers:

  • The customer asks for a person.
  • The customer expresses strong frustration or repeats the same complaint.
  • The topic is a refund dispute, safety issue or complaint about staff.
  • The agent has failed to answer twice in a row.

Then decide what the agent should say. A good handoff message is honest and brief: it confirms a person will follow up and states what the customer should expect, without inventing a wait time. If your tooling can fill in the "what happens next" detail from a real schedule, let it. Do not ask the prompt to promise "within five minutes" unless that is actually true.

Test, read transcripts, and revise

A prompt is a draft until it has met real questions. Build a small test set before launch and re-run it after every change.

  1. 01Draft the six blocks
  2. 02Test with 20 real questions
  3. 03Read the failures
  4. 04Change one thing and re-test
A simple prompt improvement loopIllustrative example

What to include in the test set:

  • Ten everyday questions the knowledge base answers well.
  • Five questions it does not answer, to check the "not sure" behaviour.
  • Three out-of-scope requests.
  • Two attempts to override the rules.

When something fails, change one thing at a time. If you edit five lines and the result improves, you will not know which one mattered. For example, if you test with 20 questions and four go wrong, fix the cause of the most common failure first and run all 20 again to check you did not break something else.

After launch, keep reading real conversations. Rated replies and flagged chats show you where the prompt is too loose or too strict, and a corrected answer is the best new example you can give the agent.

Keep the prompt maintainable

  • Prefer short rules over long paragraphs.
  • Remove rules that no longer apply.
  • Keep a dated copy of each version so you can roll back.
  • Put facts that change (prices, opening hours, policies) in the knowledge base, not in the prompt.

Doing this in AI SNS CHAT

AI SNS CHAT gives you a few places that map directly onto the steps above.

Write the persona and guardrails. Each agent has a persona and guardrails section. If you prefer not to start from a blank page, use the prompt generator, browse the prompt library, or begin from one of the 14 starter personas and edit it to match the six blocks in this article.

Choose the model. Under Agent → Settings → AI you pick the provider and model. Advanced models and model fallback need the Hobby plan or higher; on a lower plan the agent uses the standard model of the same provider. Choose the model first, then test your prompt against it, since wording can behave differently between models.

Test before customers see it. Use the Playground to chat with the agent and try your test set. Playground A/B model compare lets you run the same question against two models side by side, which is handy when deciding whether a prompt works across providers.

Give the agent the right facts. Add content under Data sources (files, text snippets, website, Q&A and more) so the "answer only from knowledge" rule has something to rely on. Keep changing facts there rather than in the prompt.

Wire up actions that the prompt mentions. If your prompt says the agent may collect an email address, enable the Collect leads action; leads appear under Activity → Leads and each one creates a contact. Contact custom attributes have descriptions that tell the agent when to ask for each field, so write them in the same plain style as your prompt. The Escalations action supports your handoff rules, and integrations need the Hobby plan or higher. Actions per agent are limited by plan (0, 5, 8, 12 or 20, rising from the lower plans to Corporate).

Handle handoff honestly. When a chat is handed to a person, the server reads the rota and adds the "what happens next" line itself, and it never invents a wait time. Helpdesk scheduling, which needs the Standard plan or higher, assigns handoffs to whoever is on shift.

Protect the agent. The per-agent security page includes prompt-injection and abuse filters, which back up the "keep your rules" line in your prompt.

Learn from real chats. Review Activity → Chat logs, where you can filter, flag and read thumbs up and down ratings. When a reply was wrong, edit it into the ideal answer to train the agent.

Roll back safely. Before a big rewrite, rely on restore points (up to 40 per agent) so you can return to a version that worked. Duplicate an agent to try a new prompt without touching the live one; it copies persona, model and training data but not conversations.

See plans and pricing and the feature overview for what each plan includes.

Conclusion

A strong support prompt is small, specific and tested. Define the role and scope, ground the agent in your knowledge, spell out escalation, shape the tone for the channel, and then read real transcripts to improve it. Treat it like a living policy document, and your agent will get more reliable with every revision.

Frequently asked questions

How long should a support system prompt be?

Long enough to cover role, scope, grounding, tone, actions and escalation, and no longer. For many support agents that means a page or less. If the prompt keeps growing, check whether facts that belong in the knowledge base have crept in.

Should I write the prompt in the customer's language?

Write it in the language you can maintain best, then add a line telling the agent to reply in the customer's language. Check a few conversations in each language you serve, since tone and politeness norms differ between markets.

Can the prompt alone stop the agent from making things up?

It helps a great deal, especially an explicit "if the knowledge does not cover it, say so" rule, but it works best together with good source content. If a question keeps failing, the usual fix is to add or correct the knowledge, not to add more prompt text.

How do I know when to change the prompt versus the knowledge base?

If the agent had the right facts but used them badly (wrong tone, no escalation, too long), change the prompt. If it lacked the facts or had outdated ones, change the knowledge. Reading a handful of failed conversations usually makes the answer obvious.

Is one prompt enough for web chat, WhatsApp and Messenger?

Often yes, if you include channel-aware style rules such as short replies and one question at a time. If a channel truly needs different behaviour, create a separate agent for it and duplicate the first to save time.

#system prompt#AI agent#customer support#prompt writing

Put this into practice with AI SNS CHAT

Build an AI agent trained on your own content and connect it to your website, LINE, WhatsApp and more. Free to start, no credit card required.