Day 3: Markdown Prompting, Content DNA and Skills, MCP Everywhere (Vapi Call, Zomato, Apify), Goose and Local Models, a Voice Agent, and the Full n8n Customer-Support Build

Template: '# Role', '# Context', '# Task', '# Constraints', '# Output format' as level-one headings, bullets under each, bold for the non-negotiables, a fenced block for any example. Why it works, as stated: training data is full of structured text, so headings and bullets are read as organisation and emphasis rather than noise; long unstructured paragraphs bury the constraint. Side-by-side on stage: the same request as a paragraph and as markdown - the markdown version followed the output format and the paragraph version dropped one constraint. Habit: keep a markdown prompt template in a notes app and fill it in.

PROCEDURE. Step 1: collect 8-10 samples of your own writing (posts, emails). Step 2: prompt 'analyse these and extract my content DNA - tone, sentence length, vocabulary, structure, what I never do - as a markdown style guide'. Step 3: test it - 'write a post about X in my content DNA' - and correct the misses back into the guide. Step 4: package it as a skill (Codex skills, Claude skills): a folder with SKILL.md holding the guide and a one-line description of when to use it, so the agent loads it only when writing as you. 'Gaslighting' demo: assert a false fact confidently and the model agrees - the lesson is that it optimises for agreement, so your style guide must state what you don't do as firmly as what you do.

A run of live demos, each one MCP server toggled on in the same Claude chat: Vapi places a real phone call and reports the conversation; Zomato searches restaurants and builds an order; Kite (Zerodha) returns a live stock quote and portfolio; Apify runs a scraper actor and returns structured rows; Indeed searches jobs by title and city. The framing: the chat window has become the operating system and MCP servers are the apps. Then the desktop layer: Goose (Block's open-source agent) runs the same servers locally with any model; Ollama and LM Studio host local models so private documents never leave the laptop - slower, smaller, but yours. Claude Design shown briefly for UI mock-ups from a prompt.

PROCEDURE in Vapi: create assistant -> system prompt (persona and the call's goal - a clinic receptionist booking appointments) -> pick voice and transcriber -> attach knowledge (a FAQ file) -> add a tool (a webhook to a booking endpoint, or a calendar) -> buy or import a phone number -> call it from a real phone on stage and book a slot. Points made: latency is the product (keep the prompt short, pick fast models); write the escalation rule ('if the caller asks for a human, say X and end the call'); log every call.

Level 1 user (chat); Level 2 prompter (frameworks, markdown, projects); Level 3 automator (n8n, Make, scheduled tasks); Level 4 agent builder (MCP, Vapi, CrewAI-style multi-agent); Level 5 generalist (ships products - Lovable, v0, Supabase). Tools named per level; Jerry mentioned as a reasoning-heavy assistant for planning and v0 for front ends. Advice: pick one project that forces the next level, not a course.

LLM: input -> predicted output, once. Agent: goal -> plan -> pick a tool -> act -> observe -> decide again, until done. Components drawn on the board: model (brain), memory (short-term context and long-term store), tools (APIs, search, code), and the loop (the orchestration that decides). The difference between a workflow (fixed steps you drew) and an agent (steps the model chooses) - and the advice that most business problems are workflows with one agent node, not autonomous agents.

Trigger nodes start the run (schedule, webhook, Gmail 'new email', form submission). Action nodes do one thing each (Gmail send, Sheets append, HTTP request, Slack post). The AI Agent node holds a model, a system prompt, optional memory, and tools it may call - it sits where a human would read and decide. Data flows as JSON items between nodes; use expressions ({{ $json.field }}) to pass fields forward. Testing: run each node alone with pinned data before wiring the whole thing. Cloud n8n (paid, easy) versus self-hosted (free, yours).

A subscription is a seat for a human in a chat window; an API key is metered access for software. Pay per token: the on-air example prices a support reply at roughly 1,500 input + 300 output tokens on a mid-tier model - about $0.002 - so a thousand tickets a month costs a couple of dollars versus a $20 seat for a person to answer them by hand. PROCEDURE: OpenAI or Anthropic platform -> create key -> set a monthly spending limit first -> paste into n8n's credential store, never into a prompt or a public repo. Model tiering: cheap models for classification, capable ones for the customer-facing draft.

PROCEDURE, node by node. 1) Gmail trigger on new mail to support@. 2) AI Agent (cheap model) classifies: order status / refund / product question / other, and extracts order number if present. 3) Switch node routes by class. 4) For product questions, a knowledge lookup (Google Docs FAQ via a Docs node, or a vector store node) feeds a second AI Agent (capable model) with the content-DNA style guide as system prompt to draft the reply. 5) Refunds go to a human: Slack post with the summary and a link. 6) Gmail 'create draft' (not send) with the reply. 7) Google Sheets append: timestamp, sender, class, action taken. Tested live with three sample emails; one misclassification fixed by adding examples to the classifier prompt. Closing rules: draft-not-send until a week of clean logs; log everything; the Sheet is your audit trail.