This is the prompt we use as a starting point when we set up an AI agent on WhatsApp for a service business (clinics, salons, spas, workshops, restaurants with reservations, etc.). It is not perfect — it is built so you can adapt it to your case in 15 minutes. Copy it, paste your data into the [BRACKETS] and test it in a sandbox before pushing it to production.
Why this prompt and not another
Most prompts circulating on YouTube are for general support chatbots. This one is optimized for three concrete things that move the needle in local businesses: responding fast, qualifying intent and booking without friction. Any message that does not fit that leaves the agent and goes to a human.
The prompt
Paste it as-is into your system (Claude, GPT-4, OpenAI Assistant API, etc.) and fill the brackets with the real information about your business.
# Identidad
Eres el asistente virtual de [NOMBRE NEGOCIO], un negocio de [SECTOR] en [CIUDAD].
Hablas en español neutro, tono cercano pero profesional, sin emojis innecesarios.
Tu rol: responder consultas, calificar la intención y agendar citas. NO inventes
servicios, precios ni horarios — si no lo sabes, lo escalas a un humano.
# Información del negocio
- Servicios: [LISTA SERVICIOS Y PRECIOS]
- Horarios: [HORARIOS]
- Ubicación: [DIRECCIÓN]
- Tiempo medio de cita: [MIN]
- Forma de pago: [EFECTIVO/TARJETA/TRANSFERENCIA]
# Cómo respondes
1. Saluda solo en el primer mensaje del día.
2. Responde en máximo 2-3 frases. Nada de párrafos largos en WhatsApp.
3. Si el cliente quiere agendar, pide nombre + servicio deseado + 2 horarios posibles.
4. Confirma siempre antes de agendar.
# Cuándo escalas a humano
- El cliente pide algo fuera de tu lista de servicios.
- Hay queja, reclamo o conflicto.
- Solicitud de descuento o presupuesto especial.
- Cualquier mensaje con palabras: urgente, emergencia, problema.
# Límites
- No prometas tiempos exactos sin confirmar agenda.
- No des consejos médicos, legales o financieros.
- Si te preguntan por la competencia, redirige al valor del negocio.How to implement it step by step
Fill the brackets with real data
Open a text editor and replace each [VARIABLE]with the exact information about your business. Do not leave any bracket empty — if it does not apply, write “not applicable” so the model treats it as explicit information.
Define your handoff logic
Decide what happens when the agent passes the conversation to a human: does it create a ticket in Slack? Does it send a WhatsApp to the owner? Does it write to Notion? Whatever it is, make sure it arrives in under 5 minutes.
Cuando detectes señal de handoff, responde con esta estructura
y marca la conversación como "needs_human":
"Entiendo. Voy a pasar tu mensaje al equipo para que te respondan
personalmente. Te escriben en menos de [X] minutos en horario laboral."Test it with 20 real conversations
Before connecting it to clients, open 20 simulated conversations in a sandbox: 10 easy ones (price inquiry, booking an appointment) and 10 hard ones (complaint, discount, after hours, trap-word). Measure in which ones the agent responded well, in which it escalated and in which it improvised.
Connect it to the WhatsApp Business Cloud API
The last step is plugging it into Meta’s Cloud API. You need a verified Business Manager, a dedicated number and template approval if you are going to do outbound. This usually takes 3–5 days.
Common mistakes we have seen
- Prompt too long. More than 1,500 words confuses the model. Be surgical.
- No escalation rules. The agent tries to solve everything and lies when it does not know.
- Scattered business information. If prices change, you have to be able to update them in a single place.
- Zero observability. Without conversation logs, you do not know what to fix.
And if you want to skip the build
If what you need is to have this running in your business without going through the learning curve, we can build the whole thing with your brand, your catalog and your CRM. The first step is a free 30-minute audit where we look at what you have today and what can be automated first.