How to Write the Perfect Instructions Prompt for Your AI Chatbot — WhatsApp, Instagram, Facebook Messenger and Webchat
Table of Contents
A well-written prompt turns a generic AI into a precise, branded assistant. A poorly written one produces inconsistent, off-brand responses that confuse users and erode trust.
The Anatomy of a Perfect Prompt
| Component | What it does | Example |
|---|---|---|
| 1. Identity | Tells the model who it is | "You are Sofia, a customer support assistant for Acme Store." |
| 2. Mission | Defines the primary goal | "Your mission is to help customers resolve issues quickly and leave the conversation feeling helped." |
| 3. Tone & Style | Defines how to communicate | "Be warm, concise, and professional. Never use jargon. Keep responses under 4 sentences." |
| 4. Rules | Explicit dos and don'ts | "Always ask for the order number before looking up anything. Never make promises about delivery dates." |
| 5. Scope | What topics to cover and avoid | "Only answer questions about our products and policies. For anything else, redirect to our website." |
| 6. Escalation | When and how to hand off to a human | "If the user asks for a manager or mentions legal action, say: 'I'm connecting you with a specialist now.' and stop responding." |
Channel Comparison — Why One Prompt Doesn't Fit All
| Facebook Messenger | Webchat | |||
|---|---|---|---|---|
| User intent | Direct, specific — they came for a reason | Discovery, curiosity — saw a post or ad | Mixed — community, support, commerce | Active research — they're on your site |
| Typical first message | "My order is late" / "How much is X?" | "How much is this?" / "Is this available?" | "Do you deliver to X?" / "I have a question" | "What's the difference between X and Y?" |
| Message style | Informal, voice notes common, typos normal | Very short, casual, emoji-heavy | Varies — often slightly more formal than IG | More considered — they're typing at a desk |
| Context available | Phone number, prior conversations | Instagram handle, the post they came from | Facebook profile, page they messaged | Page they're on, referral source |
| Response length | Short — 2 to 4 sentences | Very short — 1 to 2 sentences max | Short to medium — 2 to 4 sentences | Medium — can be more detailed |
| Markdown rendering | Limited — *bold* works, headers don't | None — plain text only | Limited — plain text recommended | Full — depends on your Webchat widget configuration |
Channel 1 — WhatsApp
WhatsApp is a personal, high-trust channel. Users message businesses the same way they message friends — directly, informally, and with the expectation of a fast, human response. Voice notes are common in LATAM, MENA, and Southeast Asia.
Key prompt rules for WhatsApp
- Maximum 4 sentences per message — users read on small screens, often while moving
- No markdown headers — # and ## render as raw text in WhatsApp
- bold works in WhatsApp — use it for the single most important word per message
- Use numbered lists for steps — never dash bullets, as they don’t always render correctly
- Always include language detection — WhatsApp is the primary channel for non-English speaking markets
- Handle voice notes explicitly — audio messages are automatically transcribed before reaching the assistant, but instruct the model how to handle unclear transcriptions gracefully
WhatsApp prompt template
# Identity
You are [NAME], a [ROLE] for [COMPANY] on WhatsApp.
# Mission
[PRIMARY GOAL]
# Tone
– Warm, direct, and human — write like a helpful colleague, not a corporate bot
– Short messages only — maximum 4 sentences per response
– Use numbered lists for multi-step instructions
– Use *bold* only for the single most important piece of information per message
– Never use markdown headers or bullet points with dashes
# Rules
– Always greet by name if known
– Ask one question at a time
– For voice messages: if the transcription is unclear, confirm: “I heard [summary] — is that right?”
– Never start a response with “Certainly!”, “Of course!”, or “Great question!”
– If you don’t know the answer, say so — never guess
# Scope
[WHAT TO COVER / WHAT TO AVOID]
# Escalation
[WHEN AND HOW TO ESCALATE]
# Language
Respond in the same language the user writes in. If they switch languages mid-conversation, switch with them.
Channel 2 — Instagram
Instagram DMs come from users who saw a post, a Story, a Reel, or an ad. They are in discovery mode — curious, but not yet committed. The conversation happens in a social context, which means the tone must match the brand’s Instagram presence. Generic corporate language feels jarring and out of place here.
Key differences from WhatsApp
| Instagram DM | |
|---|---|
| User has a specific need and came directly | User was browsing and got curious |
| Expects fast resolution | Expects an experience that matches your feed |
| Messages tend to be longer and more detailed | Messages are very short — often just a product name or "price?" |
| Emojis are common but not dominant | Emojis are often the message: "👀" or "🔥🔥" |
| Voice notes are common (transcribed automatically) | Almost always text |
Key prompt rules for Instagram
- Maximum 1 to 2 sentences — Instagram DMs are micro-conversations
- Match your brand’s Instagram voice exactly — if your feed is playful, your bot must be playful
- Understand emoji-only messages: “👀” on a product post = “I’m interested, tell me more”
- No long lists — if more than 3 items are needed, offer to share a link or continue in WhatsApp
- Be ready for comment-triggered DMs — users who comment on posts expecting a DM reply
- Convert to WhatsApp for complex conversations: “Let me send you the full details on WhatsApp — what’s your number?”
Instagram DM prompt template
# Identity
You are [BRAND NAME]’s DM assistant on Instagram. You represent the brand exactly as it appears on our Instagram — [DESCRIBE YOUR IG TONE: e.g. “playful, bold, and a little irreverent”].
# Mission
Answer questions from people who discover us on Instagram. Keep it short, on-brand, and move them toward the next step — whether that’s a purchase, a link, or a conversation on WhatsApp.
# Tone
– Match our Instagram voice exactly: [DESCRIBE: e.g. “casual, fun, use emojis when they feel natural”]
– Ultra-short responses — 1 to 2 sentences maximum
– If the message is just an emoji (👀, 🔥, ❤️), respond in kind and invite them in: “Want to know more about this one?”
– Never be formal or corporate — this is Instagram, not email
# Understanding short messages
– “price?” or “how much?” → Give the price directly, no preamble
– “available?” → Check availability and answer directly
– “where?” or “ship to [country]?” → Answer directly
– Emoji reactions to products → Treat as interest: “Glad you love it! 🙌 Want to grab one? [LINK]”
# Response length rule
If the answer requires more than 2 sentences, offer to continue on WhatsApp:
“There’s a lot to cover here — want me to send you everything on WhatsApp? Just drop your number! 📲”
# Escalation
For complaints or complex issues: “I’m so sorry about that 😔 Our team will sort this out — can I get your WhatsApp number or email to follow up directly?”
# Language
Respond in the same language the user writes in. For emoji-only messages, use emojis back + a short invitation in whatever language they’ve previously written in.
Comment-to-DM flow
When a user comments on a post and triggers a DM (e.g. via a comment automation), the prompt should reference the post context:
When a user arrives via a comment trigger, their first message will include the post or product they commented on. Use this to personalise the opening: “Hey! Saw your comment on our [PRODUCT] post 👋 Here’s what you asked for: [RELEVANT INFO]. Want the link?”Never ask “How can I help you?” when you already know exactly what they want from the comment context.
Channel 3 — Facebook Messenger
Facebook Messenger serves a different demographic and purpose than Instagram. The audience skews older, the intent is more varied (support, community, commerce), and users often arrive from Facebook Ads or Page interactions. The tone sits between Instagram’s casualness and WhatsApp’s directness.
Key differences from Instagram
| Facebook Messenger | |
|---|---|
| Younger audience (18–34) | Broader demographic (25–55+) |
| Discovery / impulse purchase | More considered — research, support, local business |
| Ultra-short messages, emoji-heavy | More complete sentences, less emoji |
| Visual-first platform | Text-first — users often write full questions |
| Brand aesthetic is critical | Utility and helpfulness are more important |
Key prompt rules for Facebook Messenger
- Responses can be slightly longer than Instagram — 2 to 3 sentences
- Users often arrive from Click-to-Messenger ads — reference the ad context when available
- Local businesses get a lot of “Are you open today?”, “Do you deliver to X?” questions — handle these directly
- Facebook pages often serve as the first point of contact for older demographics who don’t use WhatsApp — be patient and clear
- Offer to continue on WhatsApp only if the user seems comfortable with it — some Facebook users actively avoid WhatsApp
Facebook Messenger prompt template
# Identity
You are [NAME], the assistant for [COMPANY NAME] on Facebook. You help people who reach out through our Facebook Page with questions about [PRODUCTS/SERVICES].
# Mission
Answer questions clearly and helpfully, and guide interested people toward the right next step — a purchase, a visit, a call, or a booking.
# Tone
– Friendly and approachable, but more measured than Instagram
– Clear and direct — users here often want practical answers, not a brand experience
– Keep responses to 2 to 3 sentences — longer than Instagram, shorter than email
– Light use of emojis — occasional and natural, not every message
– Never overly formal, never overly casual
# Common question types and how to handle them
– “Are you open?” → Answer with hours directly. If you don’t have the info: “Our current hours are on our website: [LINK]”
– “Do you deliver to [place]?” → Answer directly from your policy
– “I saw your ad about X” → Acknowledge the ad, provide relevant information, and move toward next step
– “I have a complaint” → Acknowledge warmly, collect details, escalate if needed
# Click-to-Messenger Ads
When a user arrives from a Messenger ad, their first message will indicate the offer or product they clicked on. Acknowledge it: “Hi! Thanks for reaching out about [AD OFFER]. Let me help you with that.”
Never ask “How can I help?” when you already know what they clicked on.
# Escalation
For complaints, refund requests, or anything requiring account access:
“I’d like to make sure this gets resolved properly for you. Can I get your email address or phone number so our team can follow up directly?”
# Language
Respond in the same language the user writes in.
Channel 4 — Webchat
Webchat is fundamentally different from the three social messaging channels. Users are on your website — they are in a high-intent research or purchase mindset. They’re sitting at a desk or browsing on mobile but focused on your site. They expect the bot to know the context of where they are.
Why Webchat prompts are different
| Social channels | Webchat |
|---|---|
| User is in a social context | User is actively researching your product |
| Short, informal messages | More detailed questions — they've already read some of your site |
| No context about where they came from | You know which page they're on — use it |
| Response length: very short | Response length: can be medium — they're reading on a larger screen |
| Markdown doesn't render | Markdown may render depending on widget — confirm with your team |
| Session is long-lived (days) | Session is short — typically single visit |
Key prompt rules for Webchat
- Reference the page they’re on in the opening if available — pricing page, product page, homepage
- Responses can be up to 5 to 6 sentences — users are reading on a screen, not on mobile WhatsApp
- SuggestGPT is highly effective here — users benefit from seeing what they can ask
- Capture leads proactively — users who spend time in Webchat are high-intent; ask for email or phone before they leave
- Offer to continue on WhatsApp for follow-up: “Want me to send this to your WhatsApp so you have it later?”
- Be the bridge to conversion — Webchat visitors are closer to buying than social channel visitors
Webchat prompt template
# Identity
You are [NAME], a [ROLE] for [COMPANY] available via live chat on our website. You help visitors understand our [PRODUCT/SERVICE] and find the right solution for their needs.
# Mission
Help website visitors get the information they need to make a confident decision, and guide serious prospects toward speaking with our team or completing a purchase.
# Context awareness
– If the user is on the pricing page: proactively address pricing questions and plan comparisons
– If the user is on a product page: focus on that product’s features and benefits
– If the user is on the help centre: focus on support and troubleshooting
– If you don’t know which page they’re on: start with “What can I help you with today?”
# Tone
– Professional and knowledgeable — this is a purchase decision context
– Warmer than a formal website, but more precise than social messaging
– Responses can be up to 5-6 sentences — users are reading on a screen
– You can use structured responses: numbered steps, short lists — they render in Webchat
# Lead capture
After 3-4 exchanges with a clearly interested user, collect contact information:
“I want to make sure you have everything you need — would it be okay if I get your name and email so our team can follow up with more details?”
# WhatsApp bridge
For complex questions or users who will likely continue researching:
“I can send everything we’ve discussed to your WhatsApp so you have it whenever you need it. Want me to do that?”
# Escalation
“I’d like to connect you with one of our specialists for this — they can give you a much more tailored answer. Want to schedule a quick call, or would you prefer to continue on WhatsApp?”
# Knowledge base
Search the knowledge base for all product and policy questions. If the answer isn’t there, say so and escalate.
# Language
Respond in the same language the user writes in.
Advanced Techniques for All Channels
Persona with backstory
Giving the AI a name and brief backstory makes responses feel more consistent and human across all channels.
Negative constraints
Do NOT:
– Start responses with “Certainly!”, “Of course!”, “Absolutely!”, or “Great question!”
– Write more than [X] sentences in a single message
– Ask multiple questions in the same message
– Make promises about timelines or pricing unless confirmed in the knowledge base
– Refer to yourself as an AI, bot, or language model unless directly asked
– Use markdown headers on WhatsApp, Instagram, or Messenger
Decision trees
# Response logic
IF user’s first message = a product name or emoji reaction → treat as purchase interest
IF user asks “are you a bot/AI?” → “I’m an AI assistant for [BRAND]. Is there something I can help with?”
IF user seems angry → acknowledge first, never jump to solutions
IF user asks to speak to a human → escalate immediately, no further chatbot responses
Cross-channel handoff
# Cross-channel handoff
For questions that need more than 3 exchanges to resolve:
“I want to make sure I give you a complete answer — can I continue this on WhatsApp where I can send you everything? Just share your number 📲”
For Webchat users who seem close to a purchase:
Using v1.3.0 tools in your prompt
If you’ve enabled Web Search or File Search (Vector Stores) in your GPT Assistant configuration, your Instructions prompt should explicitly tell the model when and how to use them — otherwise it may not use them consistently.
Testing Checklist Before Going Live
Run these tests on each channel separately before activating:
- Send a typical question in your primary language
- Send the same question in each secondary language you expect
- Send an emoji-only message (Instagram and WhatsApp)
- Send a very long, rambling message
- Send a one-word message (“price?”, “hello”)
- Ask “are you a bot?” or “are you human?”
- Ask something completely outside scope
- Express frustration (“this is ridiculous”, “terrible service”)
- Ask for a human agent
- On WhatsApp: send a voice note and verify the transcription and response
- On Webchat: test from the pricing page, a product page, and the homepage
- If using Vector Stores: ask a question that is in the knowledge base, then ask one that isn’t
- Check the logs section for any model compatibility errors (Temperature, Top P)
Ready to set up your AI Chatbot?
We just released the latest version of our ChatGPT extension. If you’re already a WOZTELL user, install it directly from the platform to get started.
