Designing AI Chatbot Prompts Across WhatsApp, Instagram, Messenger and Webchat
by Jamie Chew | May 19, 2026
Table of content
WhatsApp, Instagram DM, Facebook Messenger and Webchat each carry a different user context, intent profile and conversation style. A prompt optimised for one may underperform on another. This article sets out a framework for thinking about prompt design across all four.
The Six Components of a Strong Prompt
A well-formed Instructions prompt contains six elements. Each addresses a specific failure mode that appears when it’s missing.
| 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." |
Why One Prompt Cannot Serve Four Channels
Each channel attracts a different user in a different state of mind. “How much is this?” carries different weight on Instagram (impulse curiosity from a Reel) than on Webchat (active research on a product page). The prompt has to reflect that.
WhatsApp — Personal and High-Trust
WhatsApp users message businesses the way they message friends: directly, informally, expecting a fast response. Voice notes are common in LATAM, MENA and Southeast Asia, and WhatsApp is often the primary entry point for non-English-speaking markets.
A WhatsApp prompt should enforce short responses — four sentences is a reasonable ceiling. Markdown headers do not render and should be banned outright. Bold formatting works and is best reserved for the single most important word per message. Numbered lists render more reliably than dashes. Language detection rules should be explicit, and voice-note transcriptions need a graceful handling rule — confirming an unclear transcription is better than guessing.
The most common failure on WhatsApp is porting over an email or web tone. It reads wrong instantly.
Instagram DM — Discovery, Not Resolution
Instagram users arrive from a post, a Story, a Reel or an ad. They are in discovery mode — curious but uncommitted — and the conversation sits inside a social context where corporate phrasing feels jarring.
Instagram prompts should cap responses at one or two sentences. They should include explicit handling for emoji-only messages, because “👀” on a product post is a real expression of intent. They should reference the brand’s Instagram voice directly: if the feed is playful, the assistant must be playful. Include a deliberate handoff to WhatsApp or email for anything that needs more than three exchanges to resolve.
For comment-to-DM automation flows, a context-awareness layer matters. The assistant should open with the product the user commented on, not a generic “How can I help you?” — which immediately signals an unaware bot.
Facebook Messenger — Utility Over Aesthetic
Facebook Messenger serves a broader demographic than Instagram, skewing older, with more varied intent: support, community, local commerce. Users tend to write more complete sentences, expect practical answers, and often arrive from Click-to-Messenger ads.
A Messenger prompt sits in the middle register — two to three sentences, friendly but measured, lighter on emoji than Instagram. High-frequency questions like opening hours, delivery zones and ad context should be handled explicitly.
Webchat — High Intent, Closer to Conversion
Webchat diverges most sharply from the social messengers. Users are on the brand’s website, actively researching, often comparing options. They have already read part of the site. They are closer to a purchase decision than users on any other channel.
Webchat prompts can be longer — responses of five or six sentences are acceptable. Where the platform allows, the prompt should reference page context (pricing page, product page, help centre). Lead capture should be built in proactively after three or four exchanges with a clearly interested user. And Webchat is the natural entry point to WhatsApp follow-up — capturing a number so the conversation continues after the user leaves the site.
Techniques That Generalise Across Channels
Giving the assistant a name and a brief backstory produces more consistent responses than a faceless “AI assistant.”
Negative constraints — explicit lists of phrases and behaviours to avoid, including the reflexive “Certainly!” opener — are often more effective than positive instructions.
Decision trees for predictable scenarios (angry user, request for human, “are you a bot?”) prevent the model from improvising in the moments where improvisation is most costly.
Every prompt should be tested before going live: typical questions in each supported language, emoji-only messages, one-word queries, rambling messages, frustration signals, out-of-scope queries, and explicit requests for a human agent. The test is not whether the bot responds. It’s whether the response reflects the brand.
Where to Go From Here
If you’re planning a chatbot deployment, or your existing one isn’t performing the way it should, talk to our team. We design conversational AI on the WOZTELL platform — starting with your brand, your channels and your customers.
