How to Write Chatbot Instructions and a Knowledge Base That Actually Work
Instructions are how your bot behaves; the knowledge base is what it knows. Keep them separate, write each one well, and your bot answers like someone who works there.
Updated 11 min read
Why most chatbots give vague answers
A chatbot is only as good as two things you give it: its instructions and its knowledge. Get those right and it answers like your best employee on their best day. Get them wrong and it does what disappointing chatbots do — hedge, repeat the question back, or confidently invent an answer that isn't true.
A let-down bot almost always traces back to one of two gaps. Either the instructions are too thin — "You are a helpful assistant for Acme" — so the bot has no idea how to behave, what it's supposed to cover, or what to do when it's stuck. Or the knowledge base is empty, so the bot has no real facts to draw on and falls back on generic filler.
This guide fixes both. It explains the difference between the two — instructions are how your bot behaves, knowledge is what it knows — then walks through writing each one well, setting honest boundaries, and the short test-and-refine loop that turns a mediocre bot into one customers trust. No technical background needed. If you can write a short email to a new hire, you can do this.
Instructions vs knowledge base: two different jobs
The single most useful idea in this whole guide: instructions and knowledge are two separate jobs, and mixing them up is the most common mistake new owners make.
Think of hiring someone for your front desk. On their first day you'd tell them two kinds of things. First, how to do the job: "Be warm but brief. If someone asks about pricing, walk them through the options. If you don't know something, don't guess — take their email and tell them we'll follow up." That's instructions. Second, the facts they need: your opening hours, your prices, your return policy, where the parking is. That's knowledge.
| Instructions | Knowledge base | |
|---|---|---|
| What it's for | How the bot behaves | What the bot knows |
| Contains | Role, tone, scope, what to do when stuck | Hours, prices, policies, products, FAQs |
| Length | Short — a few short paragraphs | As long as it needs to be |
| How often it changes | Rarely | Whenever a fact changes |
| Example line | "Be friendly and concise. Never quote a price you're unsure of." | "Opening hours: Mon–Fri 9am–6pm, Saturday 9am–1pm." |
When you put facts in the instructions field — "We open at 9am, a consultation is $50…" — two things go wrong. The instructions get long and muddled, so the behavior guidance gets diluted and the bot follows it less faithfully. And facts change: when your hours change you have to dig through a wall of behavior text to fix one line. Keep them apart. Instructions stay short and stable; facts live in the knowledge base where they're quick to update.
Writing instructions that work
Good instructions answer four questions for your bot, in plain language. You'll find the field labelled "Instructions for your bot" in your bot's settings.
- Who is it? Give it a role and say whose business it represents. "You are the assistant for Lumière Aesthetic Clinic, a skin and cosmetic clinic." One sentence.
- How should it sound? Set the tone in a few words plus a line of guidance. "Warm, reassuring, and brief. Use short sentences. Sound like a helpful person, not a brochure."
- What does it cover? Name the topics it's there for, so it stays on track. "You help with treatments, prices, consultations, opening hours, and how to book."
- What should it do when it's stuck? This is the one most people forget, and it's the most important. Tell the bot exactly what to do when it doesn't know or can't help. "If you don't know the answer, say so honestly and invite the visitor to leave their email so a person can follow up. Never guess a price, and never give medical advice — for any clinical question, invite the visitor to book a consultation."
Put together, a complete set of instructions reads like a short brief to a new hire:
You are the assistant for Lumière Aesthetic Clinic, a skin and cosmetic clinic.
Be warm, reassuring, and brief — short sentences, no jargon.
You help with treatments, prices, consultations, opening hours, and how to book.
If someone wants to book, or asks something you're unsure about, invite them to leave their name and email and tell them we'll get back to them quickly.
Never invent a price, and never give medical or treatment advice — for any clinical question, invite the visitor to book a consultation.
Keep it short. You have about 2,000 characters to work with — roughly four short paragraphs — and shorter is usually better. A focused half-page beats two pages of rules the bot only half-follows. Staring at a blank box? Jot down a few rough bullet points and run them through the built-in prompt optimizer (the wand button next to the instructions). It tidies rough notes into clean instructions you can then edit.
What to put in your knowledge base
The knowledge base is where the bot gets its facts. The rule of thumb: if a customer might ask it, it belongs here. Most small businesses can cover 90% of real questions with a single page of well-organized facts.
Start with the questions you already answer ten times a week. Open your email, your messages, your phone notes — the questions that keep coming back are exactly what the bot should know cold. Then add the essentials:
- Opening hours, plus holidays and exceptions.
- What you sell or offer, with prices or price ranges where you can share them.
- Location, parking, directions, and delivery or service areas.
- Policies: returns, refunds, cancellations, deposits, guarantees.
- How to book, order, or get in touch.
- The five to ten questions you're tired of answering by hand.
You can type all of this straight in, or upload it. The knowledge base accepts pasted text and PDF, DOCX, XLSX, CSV, TXT, and Markdown files, up to 5 MB each — so an existing FAQ document, a price list, or a service brochure can go in as-is, and the bot reads the text out of it. You can combine sources: paste your hours, upload your price list, paste your policies.
How to structure knowledge so the bot finds it
How you organize the knowledge matters almost as much as what's in it. A wall of text works, but a lightly structured one works better — the bot finds the right fact faster and mixes things up less.
Three habits make a real difference:
- One topic per heading. Group facts under short headings — Hours, Treatments, Prices, Consultations, Booking. The knowledge editor has a toolbar that adds headings for you, or you can type two hash marks (
##) in front of a line. Headings give the bot landmarks to navigate by. - One fact per line. "A consultation is $50 and is redeemable against any treatment booked within 30 days" on its own line is easier for the bot to quote exactly than the same fact buried inside a long paragraph.
- Write it the way a customer asks it. If people ask "do you offer laser hair removal?", have a line that says "Laser hair removal starts at $90 per session." Mirror the question, and the bot answers it cleanly.
A well-structured knowledge base looks like this:
## Opening hours
Mon–Fri: 9am–7pm
Saturday: 9am–2pm
Sunday and public holidays: closed## Consultations & prices
Consultation: $50, redeemable against any treatment booked within 30 days.
Anti-wrinkle treatments: from $250.
Dermal fillers: from $400.
Keep it current. The knowledge base keeps a version history, so you can edit freely and roll back if you change your mind. When your prices, hours, or policies change, update the knowledge base the same day. A bot quoting last year's prices is worse than no bot at all.
Handling questions it shouldn't answer
A trustworthy bot knows its limits. Two kinds of questions need a plan: the ones it shouldn't answer, and the ones it simply can't.
Shouldn't answer. Some things are too sensitive to leave to a bot — a medical symptom, a legal question, a dispute about a bill. Name these in the instructions and tell the bot to hand off rather than improvise. "For anything about a medical concern, don't give advice — invite the patient to call the clinic on the number in your knowledge." Honesty here protects your reputation.
Can't answer. For everything the bot just doesn't know, the worst possible outcome is a confident wrong answer. The fix is one line in your instructions: "If you don't know, say so plainly and offer to pass the question to a person." Then give the visitor a real way through — that's what the "Talk to a human" option is for. It's on by default: a visitor opens the chat menu, clicks it, and leaves their name and email, and the message lands in your inbox.
That said, don't overload the bot with chasing contact details. Its main job is answering questions; the handoff is the escape hatch for when it can't. Leave the door open, but let the bot do its real job first.
Test it, read the transcripts, refine
Nobody writes perfect instructions on the first try, and that's fine. The bot gets good through a short loop, not a single sitting.
Test before you launch. Every bot has a sandbox tester where you can chat with it privately, before it goes anywhere near a customer (sandbox messages cost about half the normal credits, so a thorough test run is cheap). Ask it the awkward questions: the edge cases, the things it shouldn't touch, the same question phrased three different ways. You'll spot the gaps in a few minutes.
Read real conversations. Once it's live, every chat shows up in your dashboard. This is gold. Each question the bot fumbled is a line you can add to the knowledge base; each time it drifted off-tone is a small tweak to the instructions. Skim the conversations weekly for the first month.
Patch and repeat. Saw it guess a price? Add the price to the knowledge base. Saw it ramble? Tighten the tone line. Saw a question it should escalate? Add a rule. Each fix is small, and after two or three rounds the bot handles the large majority of real questions on its own.
A worked example: before and after
Here's the difference in practice, for a real estate agency.
Before — a thin setup
Instructions: "You are a helpful assistant for a real estate agency." Knowledge base: empty.
A visitor asks: "What's your commission, and can someone show me the apartment on Mill Road this Saturday?" The bot replies with something vague and generic — "Our fees vary, please contact us for details, and viewing availability differs" — which is exactly the non-answer that sends a high-value lead straight to a competing agency.
After — a good setup
Instructions: a role (assistant for Apex Realty), a tone (professional, responsive, helpful), a scope (listings, viewings, valuations, fees, and the areas covered), and a fallback (for a specific valuation or offer, take the visitor's details and pass them to an agent; never promise a price or guarantee a viewing slot). Knowledge base: the areas covered, the fee structure, how viewings and valuations work, office hours, and the ten questions the team answers every day.
Same question, new answer: "Our sales commission is 1.5% plus tax, and we cover the whole north side of the city. I can take your name and number and have an agent confirm a viewing for Saturday — shall I do that?" That's an answer that captures a lead worth thousands.
Nothing about the second agency is more technical than the first. The owner just did the two jobs properly: told the bot how to behave, and gave it the facts. For a dozen more worked examples across different trades, see 12 chatbot examples by industry.
Common mistakes to avoid
- Facts in the instructions. Prices and hours belong in the knowledge base, not the behavior text. Keep them apart so both stay easy to manage.
- The "helpful assistant" non-instruction. "You are a helpful assistant" tells the bot nothing. Give it a role, a tone, a scope, and a fallback.
- No fallback. If you write only one rule, make it the honest "I don't know — let me pass you to a person" line.
- Dumping the whole website. A focused page of facts beats fifty pages of marketing copy.
- Stale knowledge. Update prices and hours the day they change. A wrong answer is worse than no answer.
- Over-long instructions. Past about half a page, the bot starts to lose the thread. Shorter and sharper wins.
Recap and next step
Two jobs, kept separate, and a short loop to refine them — that's the whole craft.
- Instructions tell the bot how to behave: role, tone, scope, and an honest fallback. Keep them to about half a page.
- The knowledge base holds what the bot knows: hours, prices, policies, FAQs. Structure it with headings, one fact per line, and keep it current.
- Test, read, refine. Run the awkward questions in the sandbox, read real conversations weekly, and patch the gaps.
Do those three and your bot answers like someone who actually works there. The fastest way to start is from a template that already has both written for your trade — see the plans and credits, or follow the five-minute setup guide end to end.
Let the optimizer write the first draft
Jot down a few rough notes about your bot and the built-in prompt optimizer turns them into clean instructions you can fine-tune. 50 credits free, no credit card.
Frequently asked questions
What's the difference between the chatbot instructions and the knowledge base?
Instructions are how the bot behaves — its role, tone, the topics it covers, and what to do when it's stuck. The knowledge base is what the bot knows — your hours, prices, policies, products, and FAQs. Keep them separate: instructions stay short and rarely change, while the knowledge base holds your facts and gets updated whenever something changes. Putting facts in the instructions makes them long and hard to maintain.
What should I write in the chatbot instructions?
Answer four questions in plain language: who the bot is (its role and whose business it represents), how it should sound (the tone), what it covers (the topics it's there for), and what it should do when it doesn't know (say so honestly and offer to pass the visitor to a person). Keep it to about half a page — roughly 2,000 characters. If you're stuck, write rough notes and run them through the built-in prompt optimizer.
How do I add my business information to the chatbot?
Put it in the knowledge base. Add your hours, prices, policies, and the questions customers ask most, under short headings with one fact per line. You can type it straight in or upload a document — the knowledge base accepts PDF, DOCX, XLSX, CSV, TXT, and Markdown files, up to 5 MB each, and reads the text out of them. You can combine sources: paste your hours, upload your price list, paste your policies.
Can I upload a PDF or document instead of typing everything?
Yes. The knowledge base accepts PDF, DOCX, XLSX, CSV, TXT, and Markdown files up to 5 MB each, and extracts the text for the bot to use. An existing FAQ document, price list, or service brochure can go in as-is. That said, don't upload your whole website — a focused page of clean facts works better than fifty pages of marketing copy the bot has to search through.
How do I stop the chatbot from making up answers?
Two things. First, add a clear fallback rule to the instructions: tell the bot that when it doesn't know, it should say so plainly and never guess — then offer to pass the question to a person via the "Talk to a human" option. Second, make sure the real answers are actually in the knowledge base; a bot guesses when it has nothing solid to draw on. A bot that says "I'm not sure, let me pass you to someone" is far better than one that invents a price.
How long should the instructions be, and how much knowledge does the bot need?
Keep instructions short — about half a page, or roughly 2,000 characters. A focused, sharp brief beats a long one the bot only half-follows. The knowledge base can be as long as it needs to be, but most small businesses cover around 90% of real questions with a single well-organized page of facts. Start with the questions you already answer most often and grow it from there as you read real conversations.