Agent Studio (free)

Agent Studio is a 10-step wizard inside the SocialMate app (sidebar → AI Agent). You answer a few questions about your business; it hands you a production-grade system prompt for a WhatsApp AI agent, a ready-to-paste n8n workflow, and a Claude Desktop / MCP config. It is free on every tier, including Free.

The 10 steps

1. Business. What you sell, who you sell to, your hours, your timezone (the generated scheduled workflow runs in it) and your website.
2. Language. By default the agent answers in whatever language the customer wrote in, and switches the moment they switch. Pick the languages you support, which one to fall back to when it genuinely cannot tell, and the formality of each — German du/Sie, Spanish tú/usted and Japanese keigo are different systems, so one global setting would collapse the register in at least one of them.
3. Agent. Its name, its role (support, sales, booking, orders, leads — or describe your own), its tone, and how long its replies should be.
4. Behaviour. How human it feels: the typing indicator, read receipts before replying, reply delay, whether it splits long replies into a couple of messages, reactions, and emoji. These are opinionated defaults — sensible starting points you can change, not rules and not research findings.
5. Jobs. Tick the automations it should handle — see the catalogue below. Each card tells you exactly what it needs to work (a store webhook, a calendar, a source of due invoices), so a job you switch on cannot quietly do nothing.
6. Systems. Name the systems it should reach — Google Sheets, Airtable, Notion, Postgres, MySQL, Supabase, Shopify, WooCommerce, Stripe, HubSpot, Pipedrive, Salesforce, Google Calendar, Google Drive, Slack, Gmail, Telegram, or anything else over HTTP. Each becomes a tool node wired to the AI Agent in the generated workflow, and the ones the jobs you picked actually need are suggested for you.
7. Memory. What the agent may remember: the conversation, the person, and what a photo or voice note said. On Free it can remember none of it — and the prompt says so.
8. Tools. Exactly what your agent can call, with a live count. The jobs you picked already attached what they need — pick Book an appointment and it gets the poll read-back, because an agent that can ask a question and cannot hear the answer is not an agent. This step is for the extras: seeing the photos customers send (free), confirming a queued message actually landed, working with groups. They are grouped into capability packs rather than a wall of switches — an overlapping tool list is the top documented cause of an agent reaching for the wrong tool, so the step is built to resist bloat rather than invite it.
9. Guardrails. AI disclosure (on by default), what it must escalate to a human (added to the five it always escalates), what is out of scope, out-of-hours behaviour, and any extra rules.
10. Generate. Copy the three artifacts. The prompt is a seed — a production-grade starting point that you are meant to read, edit and extend.

What comes out

1. The system prompt

Not a one-liner. The generated prompt names the exact tools your agent has — and names them correctly per surface, which matters more than it looks: under MCP the tool is whatsapp_get_ai_context, but in n8n the tool name the model sees is the node's display name (“Get Conversation Memory”). A prompt that references tools the agent cannot find is a prompt that hallucinates.

It also carries the blocks that are true of every SocialMate agent:

• The anti-ban error contract — what to do on a blocked send (rate_limit, warming, night_mode, risk_critical, duplicate_text), so the agent backs off instead of retrying in a loop against a real phone number. This is the highest-value block in the whole prompt.
• What it cannot do — no editing, no delete-for-everyone, no forwarding, no inbound typing indicator, no interactive buttons or list menus (deprecated on the WhatsApp Web protocol — send a poll instead). An agent that believes in a tool it doesn't have will invent the action and then report success.
• The honesty rules — including that it must never claim to be human.

2. The n8n workflow (JSON)

Copy the JSON, open an n8n canvas, paste. You get SocialMate Trigger → AI Agent → tool nodes already connected, with the systems you named in step 5 attached as agent tools, and the generated system prompt already in the agent's system message. Credentials are the only TODO — your SocialMate API key, your model provider, and each service you picked.

3. The Claude Desktop / MCP config

The same agent for the MCP path — a config block you drop into claude_desktop_config.json (or Cursor, Cline, any MCP client) so it can drive WhatsApp as native tools. Full walkthrough: MCP server.

The job catalogue — 21 automations

The transactional set is seeded from Meta's own WhatsApp Template Library use-case taxonomy (ORDER_CONFIRMATION, SHIPMENT_CONFIRMATION, DELIVERY_UPDATE, ORDER_DELAY, ORDER_PICK_UP, ORDER_ACTION_NEEDED, ORDER_OR_TRANSACTION_CANCEL, RETURN_CONFIRMATION, PAYMENT_DUE_REMINDER, PAYMENT_CONFIRMATION, FEEDBACK_SURVEY) — so the naming of the transactional messages isn't ours to argue about. That library is Utility-only, which is why the conversational and lifecycle jobs are SocialMate's own extension:

• Answer questions from your knowledge (FAQ deflection)
• Hand off to a human — with auto-resume when you're done
• Order status, when they ask
• Book an appointment · appointment reminders & rescheduling Pro
• Qualify and capture leads
• Cash-on-delivery confirmation Pro
• Abandoned-cart recovery Pro
• Re-engagement / win-back Pro
• Tell customers who are waiting on you Pro

Every job you tick adds its rules to the prompt, attaches the tools it needs, and emits its own workflow. Jobs your tier cannot run stay visible with a PRO chip — the wizard never hides what you'd be missing.

AI disclosure — on by default

Article 50 of the EU AI Act requires that people are informed they are interacting with an AI system, at the latest at the time of the first interaction (Art. 50(5)), and it becomes applicable on 2 August 2026. There is an exemption where it is “obvious” — but an agent deliberately built to sound human is exactly what removes it.

Agent Studio writes the disclosure line into your prompt and leaves it on. You may switch it off; even then, the generated prompt still forbids the agent from claiming to be human.

What changes on Pro

Not the generator. The agent. Generate on Free and the prompt tells it the truth:

You have no memory of past messages. Each message you receive is all you know.
Never imply you remember an earlier conversation.

That is accurate: on Free the agent genuinely cannot read the archive (Free vs Pro). Pro is what removes the line:

• Conversation memory — one GET /ai-context call hands the agent the whole thread, role-mapped and token-windowed (AI agents with real memory).
• Contact memory — the agent saves the name, email or preference it just learned, and never re-asks.
• Media memory — your model describes a photo or voice note once; SocialMate caches the description so the same file is never analysed twice.
• Scheduled & batch sends — the jobs that are reminders by nature (payment reminders, appointment reminders, cart recovery, win-back, telling customers who are waiting on you) need the smart queue to fire later.

SocialMate stores and serves what your agent learned. It never generates it — no model of ours reads your chats.

Next

• The canonical agent system prompt — the hand-authored source Agent Studio personalises.
• Recipes & example workflows
• Agent Studio — the overview page