To train a support bot on your docs, give it your help pages as knowledge, then paste a short set of instructions that limit it to those docs, set its tone, tell it what to say when the docs have no answer, and say when to hand over to a person. Below: 10 free, copy-ready prompts, one for each rule.
Training a support bot on your documentation has two halves. The knowledge half is your content: help center pages, PDFs, policies and FAQs. The instruction half is a prompt, a few paragraphs that tell the bot how to use that content. Most guides cover the first half; this one is the second. Each prompt below is written to paste as is: replace the words in [square brackets] with your own, keep the ones you need, and test them before customers see them. If you have not added your docs yet, start with how to train an AI chatbot on your documents.
Key takeaways
- A prompt is the input that steers what an AI model says.
- Most bad support answers are confident guesses.
- These are the conversations where a wrong answer costs the most.
- Two quieter failure modes: an answer translated so loosely the customer cannot find the button, and an answer that was right last month.
- The last prompt is for you, not your customers.
How do you use these prompts?
A prompt is the input that steers what an AI model says. Google Cloud describes prompt engineering as giving the model context, instructions and examples so it understands your intent (Google Cloud). For a support bot, that means one standing set of instructions that applies to every reply.
- Add your docs first. The prompts tell the bot how to use its sources; they do not replace them.
- Pick the prompts you need. Prompts 1 to 3 suit almost every support bot. Add the others for the situations your customers actually raise. The prompt picker further down shows which ones you are missing.
- Replace every [bracket]. Swap in your company name, topics, links and hours. A bracket left in place reaches customers word for word.
- Paste them as one block into your chatbot's instructions field, one prompt per paragraph, scope first.
- Test before you publish. Use Prompt 10, then read the answers against your docs.
Keep the whole block short and specific. Plain rules such as "never guess a price" are easier for a model to follow, and for you to check, than a long description of the ideal assistant. If answers still go wrong after that, our list of why AI chatbots fail covers the usual causes.
Which prompts set the bot's scope and tone?
Start here. These two lines decide what the bot talks about and how it sounds, and every later prompt builds on them.
Prompt 1: Keep the bot inside your docs (scope)
You are the support assistant for [Company]. Answer questions about [products or services] only, using only the help docs, pages and files you have been given. Do not use general knowledge to fill gaps. If a question is about something else, say politely that you can only help with [topic] and point the customer to [where they can ask instead].
Scope comes first because every other rule depends on it. Telling the model to answer only from the documents you provide, and not from its general knowledge, is one of the standard ways to reduce made-up answers (Anthropic).
Prompt 2: Set one tone of voice (tone)
Write like a helpful person on the [Company] support team: [friendly and plain / calm and formal]. Use short sentences and everyday words. Call the customer "you" and the company "we". Do not use jargon unless the customer uses it first. When a customer reports a problem, acknowledge it in one sentence before you give the fix.
Pick two or three adjectives and one concrete behaviour. "Be friendly" alone is vague; "acknowledge the problem in one sentence first" is something the model can actually follow and you can check.
Which prompts stop the bot from making things up?
Most bad support answers are confident guesses. These two prompts give the bot a safe thing to say when the docs are silent, and make every other answer traceable to a page.
Prompt 3: Say "I don't know" instead of guessing (unknown answers)
If the answer is not in your sources, say: "I don't have that in our help docs yet." Then offer [the next step: the contact form at [URL] / a handover to the team]. Never guess a price, date, feature, limit or policy, and never describe a setting that is not in the docs.
Giving the model explicit permission to admit it does not know is the first technique on Anthropic's list for reducing hallucinations (Anthropic). A clear fallback line also tells you, in the transcripts, which questions your docs are missing.
Prompt 4: Point to the page the answer came from (citing sources)
When your answer comes from a help page, end with one line: "More detail: [page title]" and its link. Use the exact names of buttons, menus and settings as they appear in the docs. If several pages apply, link the one that answers the question most directly, not all of them.
Asking for a source on each answer makes replies easy to check, for your customer and for you (Anthropic). In Dante AI, visitors only ever see pages from your website named as a source; uploaded files, text and Q&A answers are never shown to a visitor as a source (Sandbox docs), so put the pages you want linked on your site.
Which prompts handle escalation, refunds and handoff?
These are the conversations where a wrong answer costs the most. The goal is not to make the bot answer more, but to make it stop at the right moment and pass the customer on well.
Prompt 5: Know when to stop and escalate (escalation)
Do not try to solve these yourself. Offer a person straight away when the customer: asks for a human; reports a billing error or a charge they do not recognise; mentions a legal complaint, a safety issue or a data request; or is still unhappy after two of your answers. Say you are passing the conversation to the team and what happens next.
Write the triggers as a short list of situations, not a feeling such as "if things get difficult". A list is easy for the model to match and easy for you to test, one trigger at a time.
Prompt 6: Explain the refund policy without promising a refund (refund policy)
When a customer asks about a refund, explain the refund policy exactly as it is written in [policy page], and link to it. Never promise, approve, estimate or imply a refund for a specific order or account. Say that a person on the team reviews each request, and offer to pass it on with the order details.
A support bot that says "you'll get your money back" has made a promise your team may not be able to keep. This prompt lets the bot be useful (it explains the policy) while the decision stays with a person.
Prompt 7: Hand off to a human cleanly (handoff)
Before you hand over, ask for the customer's name and email if you do not have them, then write a one-line summary of the problem so the team does not have to ask again. Tell the customer that a person will reply [here in the chat / by email] and [when the team is available: hours]. Do not keep answering the same request after you hand it over.
The best handoff is the one the customer does not have to repeat. For what to look for in the tool itself, see our guide to the best AI chatbot with human handover.
Which prompts keep answers right in other languages and after product updates?
Two quieter failure modes: an answer translated so loosely the customer cannot find the button, and an answer that was right last month.
Prompt 8: Answer in the customer's language (multilingual)
Reply in the language the customer writes in. Your sources are in [language]; translate the answer, but keep product names, plan names, button labels and links exactly as they appear in the docs. If you are not sure which language the customer is using, ask in one short sentence.
Keeping interface names untranslated matters: the customer has to find that button on screen. Our guide to multilingual AI chatbots covers the rest of a multilingual setup.
Prompt 9: Handle product updates and old pages (product updates)
Our product changes often. If two sources disagree, use the one with the most recent date and say the feature may have changed. If a customer describes a feature or screen that is not in your sources, say you cannot confirm it yet and offer to pass the question to the team. Never describe a feature as coming soon unless a source says so.
A prompt cannot fix an outdated help page, so pair this one with keeping your knowledge current: re-sync the pages that changed and remove the ones you retired.
How do you test the bot with a prompt?
The last prompt is for you, not your customers. It turns the bot into its own first tester before anyone else sees it.
Prompt 10: Make the bot help you test itself (testing)
Paste this into the test chat, not the instructions: "List 15 questions a new customer would ask about [product] that your sources answer, 5 questions your sources do not answer, and 3 that should go to a person. Then answer each one as you would for a real customer."
Read every answer against your docs. The 5 unanswerable questions should get the fallback from Prompt 3, and the 3 handover questions should trigger Prompts 5 and 7. For a fuller test plan, read how to test a chatbot.
Which prompt should you use when?
Use this table to choose. Prompts 1 to 3 are the base; the rest depend on what your customers ask.
| Prompt | Purpose | When to use it | Expected answer style |
|---|---|---|---|
| 1. Scope | Keeps answers inside your docs and topics | Always, as the first paragraph | Short answer from your content, or a polite "I can only help with..." |
| 2. Tone | One consistent voice | Always | Plain sentences in your brand voice |
| 3. Unknown answers | Stops guessing | Always | A fixed fallback line plus a next step |
| 4. Citing sources | Makes answers checkable | When your docs live on public web pages | Answer, then "More detail:" with one link |
| 5. Escalation | Stops the bot at the right moment | Billing, legal, safety or upset customers | One line offering a person, no attempt to solve |
| 6. Refund policy | Explains policy without promising money | Any business that takes payments | The written policy, a link, and an offer to pass the request on |
| 7. Handoff | Passes context to your team | When a person can reply | Asks for name and email, summarises, sets expectations |
| 8. Multilingual | Replies in the customer's language | Customers write in more than one language | Translated answer with interface names unchanged |
| 9. Product updates | Avoids describing old or unreleased features | Your product or policies change often | Newest source wins, with a note when something may have changed |
| 10. Testing | Finds gaps before customers do | Before launch and after every docs change | A list of test questions and the bot's answers |
How do you set this up in Dante AI?
In Dante AI, your docs go in as knowledge: website pages it crawls, file uploads such as PDF and Word, pasted text, and question and answer pairs (Knowledge sources). The prompts go in the Behavior field, which sits on the Sandbox page while you build and under Publish, then Model, after launch; both edit the same instructions (Agent behaviour).
- Start from a preset. The Behavior field offers presets, including Customer support, that you can edit, so you can paste Prompts 1 to 9 under it and adjust.
- Stay under the limit. Instructions can be up to 6,000 characters, and short, specific ones usually work better.
- Tidy it with Refine. On paid plans, Refine instructions rewrites a rough draft into a structured prompt with a fallback for "I don't know", without inventing facts you did not give it. It stays greyed out while a [bracketed] placeholder is left in the text, which doubles as a check that you replaced them all (Refine instructions).
- Test in the Sandbox. The Sandbox is a private preview on every plan, including Free. Run Prompt 10 there; Save changes then applies your edits to the live agent straight away (Sandbox).
- Turn on Human Handover (Advanced and Pro plans). Prompts 5 and 7 work best with it on: a Talk to a human option sits under every reply, the chat lands in your Inbox, and the AI stops replying once a teammate takes it (Inbox).
One thing a prompt will not change: replies are kept to at most two short paragraphs so they stay readable in a chat window, so ask for clarity, not length.
The Free plan costs $0 and includes 100 credits a month, enough to try these prompts on your own docs. Get started free at dante-ai.com.
Sources
- Anthropic, Reduce hallucinations (allow "I don't know", restrict answers to provided documents, cite sources), read 4 November 2026.
- Anthropic, Customer support agent (staying on topic and linking to sources in support chat), read 4 November 2026.
- Google Cloud, Prompt engineering: overview and guide, read 4 November 2026.
- Dante AI Help Center, Agent behaviour, Refine instructions, Sandbox, Inbox and Knowledge sources, read 4 November 2026.
Prompt picker: which of the 10 prompts does your bot still need?
Without the interactive tool, go down the list of 10 prompts and mark each rule your current instructions already state. Copy the prompts you did not mark, starting with Prompts 1 to 3.
Frequently asked questions
What is the best prompt to train a support chatbot on documentation?
Where do I paste a support chatbot prompt?
How do I stop a support chatbot from making up answers?
Can a chatbot handle refund questions?
How long should chatbot instructions be?
Are these prompts free to use?
Ready to try this on your own site? Build your AI agent with Dante AI.
Get started free