Core Architecture

Salecium is designed with a multi-agent modular workflow. An incoming lead doesn't just talk to a generic chatbot. The system processes the conversation through distinct stages:

1. Lead Sourcing & Filtering

Leads are automatically fetched via Google Maps searches matching your criteria (e.g. minimum rating or reviews). The AI uses the analysisPrompt alongside your company context to verify compatibility.

2. Automated Outreach

Once marked qualified, Salecium delivers the custom outreach template (intro message) over WhatsApp. It skips greetings repetition and uses split messages for maximum deliverability.

3. Contextual Auto-Replies

When a lead replies, the system feeds the conversation history, customer profile data, and your full Knowledge Base into the AI using the replyPrompt. The AI qualifying rep handles questions and aims to schedule calls.

System Prompt Structure

An exceptional system prompt ensures that the AI answers accurately, stays on-brand, and triggers the correct logical sequences. We recommend formatting your prompt with four distinct blocks:

1. Identity & Persona

Define the bot's name, role, tone, and organization. E.g., "You are Sarah, a warm, professional sales representative at Hypeloop." Ensure it stays conversational, avoiding dry or overly mechanical support jargon.

2. Goals & Pipeline

Instruct the AI on what it must achieve, in order of priority. E.g., "Goal 1: Schedule a quick 2-minute phone call. Goal 2: Briefly explain product benefits. Do NOT negotiate price."

3. Rules & Guardrails

Set negative constraints. E.g., "Never quote pricing directly. If they refuse twice, end the chat gracefully. Never re-introduce yourself."

4. Instruction-Action Mapping

Teach the AI when to output Action Tags. E.g., "If they schedule a time, prepend with [BOOKED: date]. If they ask to speak to a manager, output [ESCALATE]."

Data Injection Tags

These dynamic tokens inject live data into the AI prompt layout at runtime. Ensure you write them exactly as displayed below in your system instructions.

{Knowledge Base}

Injects your entire company profile, product FAQ, pricing, and services. The AI retrieves relevant facts from this to answer queries.

{Client Name}

The name of the lead or person the AI is interacting with. Use this to personalize greetings and replies.

{Conversation History}

The ongoing history of messages. Crucial for the AI to retain thread context and continuity.

{Area}

The locality or suburb of the lead (if available in their address). Perfect for localizing offers.

{Business Name}

The name of the lead's business, allowing the AI to speak directly about their company.

Action & Logic Tags

These parameters act as automation commands. The Salecium engine scans the AI's output for these strings and immediately triggers matching operations.

[SPLIT]

Splits a long reply into multiple WhatsApp bubbles. Mimics natural human texting patterns by sending consecutive short messages.

[END]

Tells the Salecium system to stop replies. Use when the lead is disqualified, not interested, or says bye.

[ESCALATE]

Flags the chat for human attention and triggers an instant WhatsApp alert to the manager's phone.

[BOOKED: <date>]

Signals that a meeting/demo was scheduled. Updates the lead status to 'booked' and issues notifications.

[NOT_INTERESTED]

Marks the lead status as lost and politely ends the conversation context.

Conversational Rules

To build high-converting WhatsApp channels, you must follow strict formatting and design rules.

Human texting rhythms via [SPLIT]

Instead of a single massive wall of text, use [SPLIT] to send messages consecutively. WhatsApp users open short bubbles significantly more often than long paragraphs.

Continuity & repeating greetings

AI agents default to saying "Hi, how are you?" in every reply. This ruins the human feel. Instruct the prompt: "Never greet the customer again if you have already sent the intro message. Assume they know who you are."

Keep it concise

Avoid long copy or bullet points on chat replies. Limit each message block to a maximum of 2 sentences. Focus on asking qualifying questions or scheduling a call.

AI Voice Agent Guide

Voice agents require different design guidelines than chat. High-latency or long replies will lead to interruptions and poor conversation quality.

Voice Prompt Rules:

  • Keep responses extremely short: under 15 words.
  • Use simple conversational sentences (do not use brackets, lists, or HTML/markdown syntax).
  • Avoid repeating information; check context closely to pick up where the user left off.
  • Always prompt the user to continue: E.g., "Is Chennai your primary location?"

Pronunciation Dictionary:

For complex names, acronyms, or local words, configure replacements in your Voice settings to ensure the Text-to-Speech engine (e.g. Sarvam, ElevenLabs) reads them correctly. E.g., "Salecium" → "Sale-see-um".