🟢 Fundamentals & First Workflow

Build software and automations by talking to AI in natural language — you describe the result, and it handles the implementation.

Removes technical overhead: focuses on the problem, not the mechanics of each node. Speed and fast iteration.

Natural language drives the build; AI does the "heavy lifting"; iterating just means asking for a change.

An assistant that runs inside the editor: reads files, writes/edits, runs commands, and connects tools via MCP.

Understanding subscription (fixed fee, chat-based) vs. API (usage/token-based) helps avoid cost surprises.

The AI rereads the entire conversation with every message; focused sessions cost less; "one folder per project".

One folder per project + a CLAUDE.md markdown file that works as a persistent system prompt (role, context, preferences).

Defines who the AI is and how it should help throughout the project, without repeating context in every conversation.

Generate via a natural language request; editable at any time; start with an empty folder.

n8n is a visual workflow builder; an MCP server connects it to the AI, which can then read, create, test, and debug workflows.

It's the bridge that lets you build automations by chatting, without manually assembling them node by node.

Clone manually (not NPX); project-level config; API key + instance URL only in .env.

Documentation that teaches AI about n8n node types, parameters, structures, and patterns.

With skills, AI builds the right workflows — you don't need to know every node.

Installation is the same as for the MCP (only the repo changes); update CLAUDE.md; restart to register it.

A sequence of 4 tests before building: project directory, MCP connection, workflow reading, and node documentation.

Ensures the environment is 100% ready — avoids wasting time debugging setup in the middle of a build.

Four tests; common errors (partial API key, MCP won't connect, auth fails); let the AI diagnose.

The modes that control how much the AI can do on its own: ask first, edit automatically, plan, auto, bypass.

Balance speed and security — know when to let the AI act and when to approve each step.

Plan Mode plans without taking action; bypass skips approvals (dangerous); choose the mode based on context.

The AI researches nodes, asks clarifying questions, and proposes a plan for you to approve before building.

Align expectations before spending tokens building the wrong thing.

Questions (columns, extra actions); an approvable plan; pattern research.

The AI creates and validates the nodes in n8n based on the approved plan.

Know the limits: credentials, spreadsheets, and mapping are up to you.

Automatic creation + validation; clear boundaries between AI and human.

Securely connect Gmail/Google Sheets (OAuth2) and OpenAI (API key).

It's the part only you can do; understanding the flow prevents permission errors.

OAuth2 for Google; API key for OpenAI; never put secrets in chat.

Copy the error message back to the AI to diagnose and fix.

The human debugging loop is the core skill—errors quickly become fixes.

Errors as input; iteration; the AI reads, understands, and proposes a fix.

Ask for improvements through conversation—for example, changing the polling interval.

Refine without reopening the build; the workflow evolves through conversation.

Incremental changes; natural language; small, safe adjustments.

An agent that classifies emails, searches policy PDFs (vector store), and drafts replies.

It's the project that ties the whole cycle together in a real, useful use case.

RAG; text classifier; answers based on the knowledge base.

Visual checklist + test real scenarios; “green” in the editor doesn't mean the response is correct.

Prevent the agent from making up what it doesn't know—real quality, not just execution without errors.

Test as a user; the AI should refuse to make things up; green ≠ correct.

Iterate on one improvement at a time, distinguishing “works” from “works well.”

Isolated changes are easy to test and revert.

Layers; increment; measure before and after each change.

The recommended sequence of improvements: handle errors, then logic, then output quality, then performance.

Solving what matters most first keeps you from optimizing what still breaks.

Robustness before polish; performance last.

Confidence marker + code node + IF + a "let's research this and get back to you" draft when there's no answer in the knowledge base.

Ensures a decent response even when the agent doesn't know.

Confidence score; IF branch; fallback draft.

Process only emails newer than X minutes, using the "newer than" syntax in the trigger.

Avoid reprocessing old history on every run.

Time window; trigger filter; efficiency.

Detect and ignore no-reply messages, out-of-office replies, and newsletters.

Don’t waste processing (and tokens) on messages that don’t need a response.

Sender patterns; watch out for uppercase and lowercase letters (case sensitivity).

Update the agent’s system prompt and fallback draft so they sound like the brand.

Consistency of voice conveys professionalism.

Tone set in the prompt; aligned fallback; example review.

Extract the name from the email signature; use "Hello" as a fallback when there's no name.

Personalization improves the perceived quality of the response.

Name extraction; graceful fallback; dynamic greeting.

A high/medium/low badge derived from the vector store score, for internal review.

Indicates which drafts need human review before sending.

Score → range; confidence badge; triage for review.