#87: How to Vibe Code 201
Tips for getting started in Claude Code
Author’s Note: Today’s article is a sequel to #72: How to Vibe Code 101. If you haven’t read it yet, I recommend checking it out!
I’ve certainly caught the Claude Code bug. Over the past six weeks, I’ve found myself using Claude Code all day at work, as well as in the evenings for side projects. I’m consistently running up against my token usage limits and drawing on additional credits Anthropic graciously gave me.
Taking a step back, agentic coding agents like Anthropic’s Claude Code and OpenAI’s Codex allow users to build software by writing in plain English (natural language prompting). Since the February release of the Claude Opus 4.6 model, agentic coding agents have seen a step-function improvement in their ability to automate processes, reason strategically, and build legit products from a simple written request. The future truly is agentic. As we chatted about in #79: AI + Future of Work, agentic AI is the AI that is fundamentally changing how knowledge workers complete their job.
Today, I’ll dive into my vibe coding learnings as a daily user of Claude Code. For $20 (plus tax) per month, you can gain access to agentic coding agents and experience the magic of present-day technology firsthand. Not a sales pitch, just a guy who is really in awe of the ease of building today. Unfortunately, vibe coding is rarely free because of all the tokens each conversation burns when AI completes sequential workflows.
Looking back on How to Vibe Code 101, I still agree with my thought framework laid out in the article. Strong systems thinking and “starting with the end in mind” are critical to achieving your objective when working with AI.
Getting Started
Organization is a Top Priority
Before you download an agentic coding tool, you need to get organized. Comb through your folders on your computer or shared drive and make sure each file is placed in the proper folder, as well as named accurately. Your meeting notes shouldn’t be intertwined with your performance recaps.
This matters because you’ll ask Claude to complete tasks that will likely depend on where certain files live, as well as where you’d like to place the outputs of Claude’s work. For instance, if you want Claude to create a weekly performance analysis based on last week’s data, it will need to know where the information lives to complete the task.
Tip: Have a slew of random screenshots and pictures without helpful file names to tell you what the picture is? Put these files into a folder, grab the file path, and prompt Claude with “Look through the picture files in ‘X’ folder and rename each file based on the contents of the picture”. You’ll be amazed at the results.
The Interface: Terminal vs. Desktop app vs. Integrated Development Environment (IDE)
For those of you who are non-technical (like me), you probably haven’t spent much time in a terminal. Type “terminal” into your computer’s search bar and you’ll be greeted by an intimidating black screen. When I first set up Claude Code, I wasn’t too confident of my abilities to tackle using a terminal. After a week, it’s what I use 90% of the time I leverage Claude.
Once you install Claude software onto your computer, you can simply type “claude” into the ominous terminal and then you’re brought to a much friendlier user interface where you can type, and paste to your heart’s content.
Next, Anthropic launched a desktop app that allows you to access Claude Chat (a typical chatbot interface), Claude Code, and Claude Cowork all in one interface. You can view settings, manage integrations, and switch between modes here.
Finally, you can run Claude Code from within an IDE, which is essentially an interface that lets you write code (or type in English), view your file system and actual text and markdown files all in one spot. I find it helpful to review AI output and visualize how my files and folders are set up. Examples of IDEs are Visual Studio Code (what I use), JetBrains IDEs, and Cursor.
Personally, I prefer the terminal view because it gives me the maximum flexibility when using many data sources (via model context protocol) and running multiple agents at once. Desktop is much more user-friendly and intuitive, and I’ll use it for one-off tasks.
Operating within Claude Code
Use Plan Mode
I cannot urge you enough: spend as much time in “plan mode” as you can. It’ll save you time and money by using fewer tokens.
Plan mode allows you to talk back and forth with AI about what you would like to build and have it challenge your thinking. Many times, I’ll want to suggest a feature change to a product I’ve built. I’ll toggle on “plan mode” (hit Shift + Tab if you’re in the terminal or click the drop-down menu in the desktop app), type an idea, and finish the prompt with “ask me questions and challenge my thinking.”
Furthermore, Claude asks me questions about my idea, and will even request additional data points to better support why I want to take “X” action. This is critical because it forces me to be clear on what I’m looking for the AI to complete so I don’t send Claude down a rabbit hole.
The default mode in Claude is to receive a request and then execute (i.e., not plan mode). This works favorably for Anthropic because actions burn more usage tokens than back-and-forth questions. It’s a lot less expensive for Claude to write a plan than it is to take the actions within the plan. At the end of plan mode, Claude provides a formal, written plan sometimes with charts, diagrams, and visualizations. You can make edits to the plan or formally allow Claude to begin working.
Connecting Data Sources: Model Context Protocol (MCP)
One of my favorite parts of Claude Code is that I can connect various data sources to the software and have Claude develop insights from these live sources. For instance, a recent project of mine was automating performance analysis for an e-commerce funnel. I connected paid advertising data (Meta, Google Ads), search trends (via Google Search Console), website traffic information (Google Analytics 4), email performance (email service provider), and website behavior (heatmapping) to Claude through MCP.
From there, I provided context about the types of analysis I’d like to complete within the weekly performance recap. This included a mock-up of the final output, details about each data source, the personas and perspectives that Claude should consider, and benchmarks for past performance and future goals.
Then, Claude analyzed all the data sources I provided, incorporated my instructions, and produced a synthesized performance recap along with a prioritized list of priorities for brand, marketing, and technology teams to improve advertising efficiency, website performance, and conversion rates across the funnel.
In an oversimplified definition, MCPs are like APIs for AI. They allow AI to connect to external data sources and software, local files, and shared drives to stitch together context and execute requested tasks.
You can connect built-in MCPs from within the Claude Code desktop app or install them in the terminal by asking Claude for help on how to set up the MCP connection to the data source of your choice.
Note: Setting up MCPs can be “expensive”. When I say expensive, I’m referring to the context that MCP connections take up within a context window. AI currently has a limit on how much information it can recall within a single session, and that calling large external data sources through MCP takes up a large chunk of the context window as you start adding more sources (like I did in the e-commerce funnel analysis).
Be conscious of how many messages you have in a single session when leveraging MCP. Claude may begin to hallucinate or flat-out not remember context you provided earlier in the session. In this case, I recommend writing the terminal command “/compact” to summarize earlier information for Claude, freeing up context so you can continue prompting within this session.
Clarity of Thought
I’ve found Claude most effective when I provide a visualization of the process I’d like to create. For a dashboard I’m building now, I drew with a pen on paper, a front-end UX mock-up and a back-end data source map.
Now that I’m writing this out to you, I’m realizing that it likely isn’t that Claude works better with visualizations than well-written instructions. It’s that by putting pen to paper, it forced me to really think through what type of dashboard I wanted to build, how the user would interact with it, and the structure of the back end to support the various data sources.
The clearer your thought process, the stronger AI output will be.
Next week, we’ll walk through a 301 session, which will involve GitHub and pull requests (PRs), Claude skills, the importance of engineers in a vibe coding world, and plenty more tips!