So What? Setting up Claude Code for Success

Katie Robbert – 00:37

Happy Thursday. Welcome to So What, the Marketing Analytics and Insights live show. I am Katie, joined by Chris. John is somewhere in Italia eating the pasta and drinking the vino. So, fun fact: I cannot do accents to save my life, aside from my own accent, which is still sort of like a mumbles thing. Anyway, John’s on vacation; you’re stuck with the two of us.

This week, we are setting up Claude Code for success. Why should we have to do that? Claude Code is a piece of software you use to develop software. There are standards that you should be abiding by in order to set yourself up for success. These are things like the software development lifecycle and the project lifecycle, but there’s even more than that. That’s what we’re going to get into today. Chris, where would you like to start?

Christopher Penn – 01:31

Let’s start by explaining why you would use Claude Code versus another system. Katie, you’re a huge fan of Claude Cowork, which is essentially Code’s sibling. Both Cowork and Code do the same thing. They are agentic systems that you use to provide processes and project plans. They just go off, do stuff, and come back later and say, “Hey, I did the thing.”

In the five levels of AI enablement, they’re level three, where you have an agent system that goes off and does things. You can chat with it, but by and large, it will go and do things. For example, Katie, you’re working on rebuilding websites and you say, “Okay, Cowork, here’s the thing, go off and do it.” The big difference between Claude Code and Claude Cowork is what it has access to on your computer.

Cowork lives inside a virtual machine. It lives inside a little container on your computer that is locked down so that it can’t get at things you don’t want it to get at. This little Linux virtual machine has its permissions locked down. It’s a phenomenal piece of software and it’s very well engineered because it keeps the non-technical user safe from accidentally doing crazy things to their computer. However, that comes at the cost of not having a full set of access to things like command-line utilities.

Claude Code operates in a terminal. It doesn’t have a shiny UI, but it has access to your computer’s raw infrastructure. This is good and bad because you can do things very badly. If you use the “dangerously skip permissions” function, it can just delete your computer. On the flip side, it gives you full access to everything installed on your computer, so it can pick up and use other apps, especially command-line utilities.

Christopher Penn – 03:29

A real simple example: on Unix-based systems like Macs and Linux, there’s a command-line utility that’s two letters long called wc. It stands for word count. You give it an input file and it tells you how many words are in it. This sounds silly, but LLMs are very bad at counting anything. You can tell Claude Code to use the word count tool to count how many words are in a chapter you’re working on, and it will accurately say it’s 8,600.

The second reason you would use this is that Claude Code gives you more fine-grained control over your usage. Cowork is very usage-intensive. All Claude subscription plans come with a certain amount of usage in a five-hour period and a weekly period. In the Claude desktop app, you can actually see what that looks like.

On the Pro plan, which is a $20 a month plan, you can burn your weekly limit in a couple of hours of work. On the Max 20 plan, which is the $200 a month plan, we can still quickly burn the five-hour limit. Claude Code allows you to see how much any given tool is using, so you can be a little more regulated. Plus, because it’s on the command line, if you’re using tools like the HubSpot or WordPress command-line tools, that uses no AI usage because it’s a different piece of software.

Katie Robbert – 06:12

I do want to acknowledge that Claude Cowork recently added a feature this week. Now it gives you the option, similar to Code, to give it full control without it having to ask. I obviously still have “ask before acting” selected, but you can have it act without asking.

I was also testing out Claude in PowerPoint. You can now embed Claude in certain Microsoft projects. Regarding your specific language about acting dangerously, that is part of the permissions in Claude in PowerPoint.

Christopher Penn – 07:18

These sorts of things apply to all different agentic systems—OpenAI Codex, Google Anti-Gravity, Openwork, and Open Code. All these systems have usage limits because, otherwise, you would have people burning hundreds of millions of tokens. With Claude Code, you get more visibility into your usage and you can quickly understand what is soaking up memory so you can adjust accordingly. I wanted to start there because, for a lot of people, it’s not as intuitive a place to work as Cowork.

Katie Robbert – 08:03

I am one of those people. It’s interesting because I don’t find myself needing to write a lot of code, but I don’t think I realized how much code is being written in the background even while using something like Claude Cowork. We need to move past that stigma. You can do roughly the same things; it’s just a different way of approaching it in terms of how you work best.

Christopher Penn – 08:40

Exactly. Here is what we’ll do. I’m going to race through setting up a new project and then we’re going to walk through what happened step-by-step. Katie, you may want to take notes on what I need to explain because it’s going to be a lot in a very short amount of time.

Katie Robbert – 09:02

I will try to interrupt you as respectfully as possible. That said, this is a live stream, and if people want to ask more in-depth questions, they can join our free Slack community at TrustInsights.ai/analyticsformarketers. If you just need someone to do it for me, go to TrustInsights.ai/contact.

Christopher Penn – 09:36

Step one: start a new folder and open a terminal in that folder. I’ve got a folder called “Live Stream Claude Code Setup” on my desktop. It is completely empty and ready to go. First, I’m going to use a custom-built piece of software called Claude Setup. It asks what I am doing—Python, React, WordPress, fiction, business? For today, I’m going to choose 11, 12, and 13.

Katie Robbert – 10:14

I’m going to pause you there for a hot second. Chris, you mentioned this is a custom piece of code you built. One of the things it does is create the project structure—your folders, agents, skills, and plugins. We’ve covered setting up Claude Desktop on previous live streams on our YouTube playlist. Instead of creating all these folders manually, you built a script that just does it. This is specific to the kinds of projects you typically undergo, so I want to make sure people aren’t looking for this specific menu when they start a project. You built this.

Christopher Penn – 11:05

Exactly. This is something I built. It copies in tons of pre-baked material. Next, I’m going to install the memory system. I’m telling Graphify to install itself, then we’ll do a Graphify update. Now I’m ready to actually start Claude. I turn on Claude, and here we are.

When we talk about setting up Claude Code for success, that is what the final polished version looks like. In my folders, I have my reference documents, agents, plugins, and skills. You can see things like the “fact check skill,” which is a free giveaway we have in the Trust Insights Academy. It also has references. For example, when I’m doing slides for a conference, I have a massive document about good slide design and a checklist of 40 things every presentation must follow so I’m presenting in alignment with neuroscience.

Katie Robbert – 12:55

Okay, so first you opened a new folder and a terminal. If you didn’t have your pre-baked script, you could set this up manually by creating folders and pulling in agents. Can you ask Claude Code, “What do I need to do to get started?” Will it give you a list?

Christopher Penn – 13:24

No, it won’t because it doesn’t know you. This goes to the heart of today’s episode: you have to have your own governance and structure mentally before you even start using the tool. A big part of what I do—and I stole this 100% from you, Katie—is to have a repeatable process in place so you aren’t reinventing the wheel or forgetting things.

I evolved to this point by asking, “What do I keep doing over and over again?” I realized I was copy-pasting the same structures, which was stupid. So I went to Claude Code and said, “I keep copy-pasting the same thing. Help me design a system that will do it for me.” That is how it came up with the menu items.

There is a master repository on my computer where all the prep materials are drawn from. It contains skills, references, plugins, and the setup script. For example, I have a PPTX skill for making PowerPoints. That belongs in my “Office Tasks” and “Trust Insights Tasks,” but not in my “Python Tasks.” Governance and process are two things you cannot escape, even with AI.

Christopher Penn – 16:16

If anything, it’s even more rigid now. In the Documents folder, I have all the different rules files and checklists I’ve built over time. One of the biggest mistakes people make with AI is assuming the model knows what it’s doing in their specific context. AI knows probabilities. But if you give it the right information upfront, you don’t have to rely on its background knowledge.

For all major projects, I use a deep research process with the Trust Insights CASINO framework to build reports. I typically do this in three different tools—Gemini, Perplexity, and Alibaba Qwen—and then merge the reports into a single rules file. It includes first principles, best practices, anti-patterns, and design patterns.

Alibaba published a paper at the end of 2025 stating that if you want phenomenal performance out of AI, you should always give it a validation checklist. For my song lyrics, for example, the AI goes through a checklist to ensure multi-level mnemonic stacking and “earworm dosages.”

Katie Robbert – 19:34

You are a strange human, Chris Penn. But I like this because it highlights that setting Claude Code up for success means the human still has a heck of a lot of work to do before even entering the system. That’s true of all software, but AI highlights where you’ve skipped steps because it sounds authoritative even when it’s making mistakes.

Christopher Penn – 20:12

Exactly. My lyrics checklist was derived from 35 different academic papers that I synthesized using NotebookLM. You start with deep research, build a compendium of best practices, and distill that into a checklist. As you work, use the checklist to validate the output. If it fails, look at the reference guide to see what to do differently. It’s no different than training a human employee.

Katie Robbert – 21:11

In software development, you should have those first principles—the best practices of your specific organization. It covers when to go in or out of scope, when to ask for help, or when to branch code. When coding with AI, you want that mindset so you avoid “vibe coding,” where you just see what happens. You’ll burn through your usage trying to iterate when you could have just done the prep work upfront.

Christopher Penn – 22:27

You end up doing a lot of rework otherwise. Anthropic says that when you’re doing projects, you should generally create a file called claude.md. This file contains the rules for that specific project. Their advice is to document what Claude has done wrong in the past and tell it what to do instead.

After my script runs, I run an installation for a piece of software called Graphify. It is a free, open-source knowledge graph for LLMs and humans. It creates a knowledge graph of your codebase so the AI can understand what is in your project. Without a knowledge graph, if you ask Claude to fix a function, it starts doing “find and replace” all over your code, which hits your hourly limit fast. With a knowledge graph, it can quickly see which files use that specific function.

Katie Robbert – 24:30

I have a quick question. You’re using the terminal, which is what you’re comfortable with. You can also access Claude Code through the desktop and the web version. What is the difference?

Christopher Penn – 25:01

The web browser version is substantially limited. It operates in its own virtual machine and you need to connect it to a GitHub repository to get value out of it. The desktop version integrates with connectors, which is handy, but it doesn’t give you direct access to your machine the way the terminal does.

However, the web and desktop versions are good for “teleporting.” You can start work on your machine and pick it up on your phone. If you want to ensure you have no work-life balance whatsoever, this is the best way to do it. But the terminal version remains the most powerful because it is directly on your machine.

Katie Robbert – 26:41

It’s as if before you even get started, you should run through the 5P Framework to figure out the purpose, people, process, platform, and performance. Pick your platform based on that, because the web, desktop, and terminal versions offer very different experiences. Always start with “why” before you open a terminal and play around, because playing around will cost you.

Christopher Penn – 27:23

Graphify explores the semantics and topics within a project. In this mystery thriller book I’m working on, there are two webs: one for the code that measures writing fluency and one for characters, settings, and props. By installing a knowledge graph, you can reduce your token usage by about 90% because the AI isn’t inefficiently reading every file over and over. It just finds the specific thing it needs.

Katie Robbert – 29:15

Say again what they need to install?

Christopher Penn – 29:19

It’s called Graphify. The easiest way is to copy the URL into Claude Code and say, “Install this for me, please.” We will put that URL in our free Slack group for our users.

The first thing I do is have the AI build its first knowledge graph. It analyzes the current state and then updates itself at the end of every major turn. This keeps the project on the rails. AI coding agents have a bad habit of writing “dead code”—code that isn’t connected to anything. In a knowledge graph, you can see those clusters and tell Claude to delete them.

Christopher Penn – 31:22

The next step is to create three documents in sequence: the PRD, the Spec, and the Work Plan. The PRD is the Product Requirements Document (the “why”), the Spec is the Technical Spec (the “what”), and the Work Plan is the implementation plan (the “how”). This applies to everything. If you’re writing a marketing book, you still need the why, what, and how.

Katie Robbert – 32:20

Where is the measurement plan?

Christopher Penn – 32:23

The measurement plan belongs at the end of the PRD and the Spec. The PRD has outcomes and the Spec has KPIs and milestones.

There is also a plugin marketplace inside Claude Code. If you install only one, make it the “Superpowers” plugin. It was built by Jesse Obra and Anthropic eventually acquired it. It contains functions for brainstorming, sub-agent driven development, and parallel agents. It’s very token-efficient.

Katie Robbert – 34:04

I bring that up because you should research who created a plugin so you don’t accidentally open your company ecosystem to vulnerabilities.

Christopher Penn – 34:43

A few other success-adjacent commands: usage gives you your usage chart, and context tells you what is occupying Claude’s context window. If you have MCPs (Model Context Protocol) turned on that you don’t need, turn them off. They consume tokens in every single chat turn, and those savings add up fast.

Recently, Anthropic introduced effort. You can crank the dial up or down. If you set it to low, it responds faster but might give nonsensical results because it isn’t “thinking.” If you crank it to max, it does great work but burns through your quota.

You can also change the model. They often recommend Opus, but that burns quota. I recommend “Plan Mode.” In Plan Mode, it uses Opus to write a thoughtful plan, and in “Execution Mode,” it uses Sonnet, which is much cheaper. You can toggle these with Shift+Tab.

Christopher Penn – 39:19

Our next step is to brainstorm using the 5P Framework. I use a free tool called Typewhisper. It’s a Mac and Windows app that uses an on-device model, so your data never leaves your computer.

For a new application that turns descriptions into ASCII art: the Purpose is to make ASCII art; the People are technical nerds; the Process involves creating a PRD, Spec, and Work Plan; the Platform is Python 3.12; and Performance requires 100% test coverage.

It takes time to build these libraries with deep research, but it is worth it for the consistency, quality, and reduced hallucination. Anything you do more than three times, turn it into a skill.

Katie Robbert – 43:21

Do you have templates for the PRD, Spec, and Work Plan in your preparation files?

Christopher Penn – 43:35

No, I usually dictate them because they vary wildly. A PRD for a Python project is very different than one for a novel. However, if your organization has specific coding or brand standards, those should absolutely be in your Documents folder as templates.

Katie Robbert – 44:30

We have a question: “Do you have advice about folder structure management? For example, if you copy an existing structure, does everything carry over?”

Christopher Penn – 47:52

The answer is yes, but it depends on how you work. Your folder structure should represent your workflow. Katie and I organize information very differently, and that’s okay.

To answer the second part: have Claude help you with a setup script. The magic word you want to use is idempotent. This means the state doesn’t change when you rerun the software. If I run the same setup script 100 times, it shouldn’t break what’s already there; it should just add anything new.

Katie Robbert – 49:37

I’ve had Claude clean up my desktop folders before. When I got lazy with my structures, I gave it a master folder and told it to organize everything by project. You can certainly have Claude help you with that, but I’d reserve that for the end of the week if you have “use it or lose it” usage left.

Christopher Penn – 50:13

That’s it for setting up Claude Code for success. You’ll notice we spent very little time actually inside Claude Code because most of the setup—the research, the checklists, the knowledge graph—occurs outside of it. If you do these things, you will increase your likelihood of success. We will see you all next time.