Write instructions

What it is

Instructions are standing rules your chatbot follows in every reply: how it should sound, what it must never say, how it should lay an answer out. They are a short list of separate items, each up to 200 characters — not one long brief.

The distinction that matters most: instructions steer how the chatbot answers; knowledge decides what it can answer. An instruction cannot teach your chatbot a fact. Putting your refund policy in an instruction and expecting the chatbot to quote it accurately will not work — that belongs in Knowledge. Instructions are the house style guide sitting on top of it.

When you would use it

Come here after you have loaded knowledge and the answers are factually right but the delivery is wrong:

• The tone is off — too stiff, too chatty, wrong level of formality for your market.
• Answers are walls of text when your customers want three bullets.
• The chatbot volunteers things you would rather it did not: guesses at prices, opinions on competitors, medical or legal interpretation.
• It answers in the wrong language, or switches language mid-conversation.
• It needs a consistent sign-off, or should always point at a specific next step.

If the chatbot is getting facts wrong, an instruction is the wrong tool. Fix the knowledge instead — see Chatbot will not answer.

Where to find it

Open Dashboard → Chatbots → your chatbot → Settings → Instructions (?tab=instructions).

The same list appears as the Custom Instructions step when you first create a chatbot, so anything you set in the wizard is what you will find here later. See Create your first chatbot.

Steps

1. Open the chatbot and go to Settings → Instructions.
2. Read the section description — it tells you how many instructions your plan allows on this chatbot.
3. Select Add Instruction. The button shows your current count against the maximum, so you always know how many slots are left.
4. Type one rule. One rule per item — do not stack three behaviours into a single line.
5. Watch the character count. An item over 200 characters is rejected with "Instruction must be 200 characters or less" and Save stays disabled until you shorten it.
6. Add further items while your plan has slots.
7. Select Save. Unsaved edits show Save and Reset on the section, so you can back out of a change you do not like.
8. Open Test Chatbot and ask something that should trigger the rule. Check the wording changed the way you expected. See Try the test chat.
9. Delete anything you are not ready to enforce. An instruction you are unsure about is worse than no instruction.

Writing instructions that actually change the answer

The single biggest failure is vagueness. "Be helpful" changes nothing — the chatbot is already trying to be helpful. A rule works when the chatbot can tell, while writing, whether it is obeying it.

Be concrete about the behaviour.

Weak	Strong
Be professional	Always respond in a professional, friendly tone; never use slang or emoji
Keep it short	Use short paragraphs and bullet points; never write more than five sentences before a break
Do not guess prices	Never invent a price. If the price is not in your knowledge, say you will check and offer to take their email
Be careful about medical stuff	Never interpret symptoms or suggest treatment. Say a clinician must advise, and offer to book a call

Write the prohibition as well as the preference. "Never" rules are followed more reliably than "prefer" rules, because there is no judgement call in the middle.

Name the format when you care about it. "When explaining steps, number them and use bold for screen names" produces consistent output; "make it clear" does not.

Address the reader. "Address the user directly using 'you' and avoid corporate jargon" is a single instruction that fixes a whole category of stilted replies.

A worked set for a small support chatbot

Four items, each doing a distinct job:

1. Always answer in the language the visitor used. Never switch languages mid-conversation.
2. Never invent a price, a delivery date, or a policy. If it is not in your knowledge, say you will check.
3. Use short paragraphs and bullet points. Keep answers under 120 words unless the visitor asks for detail.
4. When a visitor is upset or asks for a person, apologise once, do not defend the company, and offer to pass it to the team.

Notice that the fourth sets the tone around an escalation but does not perform one. Actually handing the conversation over is a Call Action — see Hand off to a human.

What instructions cannot do

Worth knowing before you write something that will quietly not work:

• They do not add facts. The chatbot still answers only from your knowledge base. An instruction saying "our warranty is two years" is not a reliable source of that fact — put it in knowledge.
• They do not run anything. Instructions never trigger an action, send an email, or call an API. That is the chatbot's Actions tab. See What are Call actions?.
• They do not override safety or product behaviour. If retrieval finds nothing relevant, the chatbot says so — an instruction telling it to always give an answer will not change that.
• They are per chatbot. Each chatbot has its own list. Rules do not propagate between chatbots in the same workspace.
• They are not the welcome or fallback message. Those are separate fields — see Chatbot messages and language.

Instructions and Call actions together

The two are designed to be used as a pair. Actions give the chatbot capabilities; instructions set the policy around using them.

A typical pairing: enable a handoff action, then add the instruction "When a visitor asks for a refund, do not quote the policy — offer to pass the case to a person." The action provides the escape hatch; the instruction makes sure the chatbot takes it instead of improvising.

Keep the trigger logic in the action's own When should the AI use this? description, not in an instruction. That description is what the model reads when deciding whether to fire the action, and it is far more effective there.

What you will see

The Instructions tab is a numbered list editor. Each row is a text field with a character counter and a remove control. Add Instruction shows your count against your plan maximum and disables at the cap. Call actions live on the chatbot's separate Actions tab.

If your plan allows fewer instructions than a higher plan would, a Want more instructions? notice appears with an upgrade prompt. It is shown only when a higher plan genuinely offers more, so you never see a pointless nudge on the top plan.

Limits and plan notes

• Each item is capped at 200 characters.
• The number of items is capped by your plan: Free 1, Starter 3, Standard 8, Pro 12, Agency 20 per chatbot. See Instruction limits.
• You need permission to edit that chatbot's settings; a Viewer cannot save changes. See Roles and permissions.
• Instruction slots are per chatbot, not per teammate, and are separate from your Call Action allowance and your knowledge allowance.
• Editing instructions does not re-train your chatbot or re-process your knowledge. The change applies to the next reply.

Common problems

I cannot add another item.

You are at your plan's cap and Add Instruction is disabled. Combine two short rules into one item, delete one you no longer need, or upgrade. See Upgrade or downgrade.

A long paragraph is rejected.

Items are capped at 200 characters each. Split it across items if you have slots, or cut it down. This field is not a document — if you are trying to paste a policy, it belongs in knowledge.

The chatbot ignores the rule.

Three checks, in order. Did you Save (the section shows Save / Reset while there are unsaved edits)? Is the rule concrete enough that the chatbot could tell it was breaking it? And does your knowledge contradict it — a document that states the opposite will usually win, because the chatbot is grounded in your content.

Two rules fight each other.

"Keep answers under 100 words" and "always explain the full process step by step" cannot both be satisfied. Pick the one that matters and delete the other; a contradiction makes behaviour inconsistent rather than averaged.

I changed the tone and now answers feel robotic.

Over-constraining does that. Try removing your format rules and keeping only the tone and prohibition rules — those two categories deliver most of the benefit with least of the damage.