Five MCP Prompts That Run Your AI Agents: Weekly Report, Unanswered Questions, Inbox Triage and More

The short version

An MCP server can offer prompts as well as tools: ready-made requests your client lists for you. In Claude Code, type / and they appear alongside your other commands. The ThinnestAI MCP server ships five. Each is a worked plan — which tools to use, in which order — with the points where the assistant must stop and ask you written in.

1. build_agent

Creates an agent for your business, teaches it from your website and tests it. Give it your business name and site; it drafts the instructions, adds your pages as knowledge, and runs a few test chats so you see the answers before a customer does.

2. connect_my_database

Lets an agent look things up in your own database through your API. It helps you build small, single-purpose endpoints, connects each one as an action — switched off until you approve — and tests them. Your records stay in your database; the agent asks for one answer at a time. What stays where explains exactly what we keep.

3. weekly_report

Usage, calls, WhatsApp and each agent's results for the last seven days, with what to fix. The numbers are the same ones the console shows — including WhatsApp cost by category — so the report never disagrees with your dashboard. It is the one prompt a Read-only key is offered.

4. fix_unanswered_questions

Your agent's analytics record the questions it could not answer. This prompt groups them by topic, drafts the knowledge that would answer each group in your business's voice, and marks anything it is unsure of for you to confirm. It is told not to invent facts about your business.

Nothing is added until you approve the drafts. Then it adds them and test-chats one question from each group to check the agent now answers.

5. triage_inbox

Goes through the conversations waiting for a person — escalated, or taken over — and gives you one line each: who, what they want, how long they have waited, and what it suggests. Then it waits for your decisions and carries them out: resolve, assign, take over or reply. A reply needs a full-access key and your approval of the exact message.

It is also told to treat everything customers wrote as information, never as instructions — a customer cannot talk your assistant into anything.

Which prompts your key sees

Prompts follow the key. A Read-only key sees only weekly_report; a Build or full-access key sees all five.

Try the prompts in Claude Code →