Getting Started: Claude Code for Economists with Paul Goldsmith-Pinkham | Markus Academy | Ep. 162-1

Introduction to Cloud Code for Applied Economists

Overview of the Miniseries

• The miniseries focuses on using LLMs (Large Language Models) for empirical research, featuring Paul Goldsmith Pinkham.

• Previous sessions included Ben Golop discussing theoretical economics and practical applications with tools like Kursa and refine.inc for improving academic papers.

Today's Focus

• Paul will present on Claude code and its application in applied economic research, emphasizing the rapid advancements since December 2025.

Motivation Behind Using LLMs

Benefits of LLM Tools

• LLM tools significantly accelerate the process from idea generation to empirical results, enhancing productivity in coding and data analysis.

• While raw AI outputs may lack depth without human insight, the speed of execution is a notable advantage for empirical researchers.

Encouraging Reproducibility

• Utilizing these tools promotes better reproducibility frameworks by tracking processes automatically during data analysis.

Understanding Limitations and Capabilities

Importance of Familiarity with LLM Functionality

• Even skeptics should understand how LLM works to recognize its limitations, which is crucial for roles such as referees or co-authors.

Audience Engagement

• The presentation aims to cater to both newcomers and those familiar with Claude code, providing foundational concepts before diving deeper into practical applications.

Upcoming Video Content

Data Analysis Tasks

• Future videos will cover simple data analysis tasks including web data retrieval and visualization related to home ownership in the U.S.

Advanced Data Scraping Techniques

• A session will focus on scraping complex unstructured text data from Edgar, a database containing public firm filings.

Working with Big Data Sets

Enhancing Skills with Large Datasets

• Claude and similar LLM technologies can improve handling big datasets through better structuring and parallel processing capabilities on high-performance clusters.

Writing Assistance Using LLM Tools

Constructing Task Lists from Referee Responses

• Paul plans to demonstrate how Claude can assist in organizing tasks based on referee feedback, showcasing its utility even for those less experienced in coding.

Introduction to Claude Code

Overview of the Video Content

• The video discusses customization techniques for Claude, focusing on best practices and Git workflows.

• It introduces Claude Code as a primary topic, comparing it with other tools and providing installation guidelines.

• Pricing information is shared, noting that using Claude Code is comparable to other AI services if you already pay for Claude.

• The concept of the context window is introduced, emphasizing its importance in discussions about data privacy.

• A brief overview of command line enhancements is provided, leading into a summary of what will be covered.

What is Claude Code?

• Claude Code can be viewed as an LLM (Large Language Model) operating locally on your computer while still accessing the internet.

• Users can input commands in natural language instead of needing to know complex scripting languages; this simplifies coding tasks significantly.

• The analogy of having an RA (Research Assistant) illustrates how Claude can read code, run scripts, and build projects through user interaction.

Benefits of Using Claude

• Unlike traditional email exchanges with experts where delays may occur due to lack of access to necessary files, Claude can instantly process large documents for summarization or analysis.

• Kyle Jensen's "pathway to enlightenment" metaphor describes different levels of AI usage among users—from basic browser interactions with ChatGPT to more advanced IDE integrations.

Integration with Development Environments

Levels of AI Interaction

• Many users are at level zero or one—using ChatGPT in browsers or copying code snippets into editors after generating them via AI assistance.

• Level one involves using Integrated Development Environments (IDEs), such as VS Code or Cursor, which allow LLM integration for enhanced coding support.

Tools and Features

• Microsoft Copilot within VS Code exemplifies inline completions that assist developers by integrating LLM capabilities directly into their workflow.

• Distinctions between Microsoft Copilot (for Word processing tasks) and GitHub Copilot (for coding assistance within IDE environments), both leveraging LLM technology.

Future Directions

• As users progress beyond basic interactions towards agentic AI capabilities, they will engage with systems that autonomously perform tasks like reading and executing commands without constant user input.

AI Agents and Coding Tools

The Role of AI in Coding Environments

• Discussion on how AI can operate autonomously, executing tasks without needing constant user input. This includes running its own code and searching the web.

• Introduction to Cursor as a powerful agentic IDE that allows users to command LLMs (Large Language Models) to perform various tasks seamlessly across the interface.

• Explanation of task delegation where users assign specific roles to agents, allowing them to figure out necessary steps for completing high-level tasks like updating a CV with papers from a website.

Advancements in Dedicated Coding Agents

• Overview of dedicated coding agents that are becoming increasingly popular, highlighting tools like Claude Code and Co-work as examples of open-source software designed for this purpose.

• Description of capabilities such as reading/writing files, using tools, executing commands, and creating plans within these environments.

Understanding Different Levels of AI Integration

• Clarification on the distinction between different levels of AI integration in coding environments; Level 2 involves employing multiple agents within an editor while Level 3 allows for more advanced interactions.

• Mention of various open-source versions available that utilize LLM technology, emphasizing their role as programs harnessing cloud-based LLM capabilities.

Enhancing Agent Performance

• Discussion on improving agent performance through updates related to skills or memory banks—though not true memory—allowing agents to track tasks effectively.

• Insight into how boundaries between different levels (2 and 3) may blur due to similar functionalities offered by tools like Claude Code.

Future Directions and Capabilities

• Exploration of future possibilities where agents could work independently over extended periods, potentially leading to significant outputs like writing entire papers autonomously.

• Acknowledgment that all discussed systems ultimately rely on the same underlying LLM technology despite variations in toolsets and functionalities.

Key Tools in AI-Assisted Development

• Listing key tools such as Claude Code, Claude Co-work, Cursor, Copilot, Zai among others that represent varying levels of integration with coding environments.

• Emphasis on understanding how these tools function at different levels—from basic assistance (Level 1 & 2) up to more autonomous operations (Level 3).

This structured overview captures essential insights from the transcript regarding advancements in AI-assisted coding environments while providing timestamps for easy reference.

Understanding Claude Code and Co-Work

Introduction to Claude Code

• The speaker humorously reflects on feeling nostalgic about computing, likening it to being back in a lab using MS DOS. The main point is that with Claude Code, users don't need extensive command-line knowledge as it automates many tasks.

Terminal Usage and Shell Commands

• Familiarity with the terminal enhances user experience; it allows access to the entire file system and execution of shell commands, which are essential for running scripts and managing projects effectively.

Comparison of Claude Code and Co-Work

• The speaker contrasts Claude Code with Co-Work, suggesting that while Co-Work is user-friendly for those uncomfortable with command lines, it's not suitable for full-time coding due to its limitations.

Features of Co-Work

• Co-Work operates like a web browser on your computer but has restricted capabilities. It can perform basic tasks without full autonomy compared to Claude Code, making it safer but less powerful.

Limitations of Sandbox Environment

• The sandboxing feature in Co-Work limits internet access and functionality. While this reduces risk, it also restricts what users can do compared to the more open environment provided by Claude Code.

Integration within IDEs

• Using Claude within an Integrated Development Environment (IDE) allows some local operations but still lacks complete access to bash commands. This creates a balance between usability and functionality.

Operational Differences Between Tools

• The speaker emphasizes that while both tools serve similar purposes, Claude Code offers a more comprehensive operational experience suited for research workflows compared to the limited scope of Co-Work.

Internet Access Management

• Users must manage permissions carefully in Co-Work since it cannot run scripts autonomously or scrape data from the internet without explicit allowances.

Installation Process Overview

• To install Claude Code or Co-Work, users can easily follow straightforward instructions available online or through terminal commands. Emphasis is placed on user-friendliness during installation processes.

Cost Structure of Claude Chat

Overview of Subscription Levels

• There are three subscription levels for Claude: Pro, Max, and Max 20x. The Pro level is recommended for light users to explore the platform's capabilities.

• The speaker personally uses the Max subscription due to high usage but suggests starting with either the $20 or $100 options for most users.

• The subscriptions are described as heavily cross-subsidized, indicating that they provide good value relative to their cost.

Understanding Context Window

• A key concept discussed is the "context window," which refers to how interactions with the LLM (Large Language Model) are structured and processed.

• Users typically see only a portion of the interaction (the prompt and response), while much of the internal processing remains hidden from view.

Interaction Dynamics

• Every user query sends back all previous messages in a conversation, creating a cached message system that retains context throughout interactions.

• This leads to potential performance degradation over longer conversations, as irrelevant information may accumulate.

Importance of Context Engineering

• The idea of "context engineering" is introduced, emphasizing that precise prompts lead to better performance from the LLM.

• As conversations lengthen, it becomes crucial to summarize and condense information effectively to maintain clarity and relevance.

Strategies for Effective Use

• Users should break down tasks into actionable steps and document findings systematically in a research file before initiating new conversations based on those insights.

• Recent improvements in models like Claude Code have made them more adept at managing context without requiring extensive manual summarization by users.

Memory Management Techniques

• Users can guide what information gets retained or forgotten by using commands like "compact," allowing them to manage memory within their interactions effectively.

Understanding Nonlinear Programming and Context Windows

Key Concepts in Nonlinear Programming

• The discussion emphasizes the importance of recalling concepts related to nonlinear programming, suggesting that a focused approach aids memory retention.

• The "context window" is described as a shared resource; being prepared enhances efficiency, while lack of focus leads to wasted time.

Privacy Considerations with AI Tools

• A warning about privacy issues when using AI tools is presented, noting that data interactions occur during tool calls and conversations are processed through APIs.

• The speaker advises caution regarding data handling by AI companies, highlighting varying legal perspectives on data usage and the need for careful management of sensitive information.

Data Security Best Practices

• Recommendations include isolating sensitive data (e.g., PII or IRB-related information) from general use cases to mitigate risks associated with cloud services.

• Users should avoid pasting sensitive information like API keys into public or unsecured environments; tools often provide warnings against such actions.

Enhancing Terminal Usability

• Suggestions for improving terminal aesthetics and functionality are provided, including tools like Ghosty or Zealage for better user experience.

• An explanation of shell commands clarifies their role in executing tasks within Unix/Linux systems, contrasting them with Windows command structures.

Practical Tips for Using AI Effectively

• To maximize productivity with AI tools, users should be specific in their requests—providing clear instructions similar to guiding a research assistant.

• Iteration is encouraged; if an AI tool begins to deviate from expected outputs, users can interrupt processes and reset context effectively.

Trusting AI Outputs

• While acknowledging the capabilities of AI programs in generating code or analyses, users are reminded to verify results due to potential errors.

• Emphasizing the importance of oversight when working with automated systems parallels managing human assistants—trust but verify remains crucial.

Cloud Code vs. Cloud Co-Work: Understanding the Differences

Overview of Cloud Code and Its Advantages

• The advantage of using cloud code over sandboxed environments like cloud co-work is that it integrates seamlessly with existing tools on your computer, allowing for flexibility in programming languages such as MATLAB or Python.

Types of Agentic Tools Discussed

• Different types of agentic tools were covered, including:

• Cloud Code: Operates in the terminal with full access to system resources.

• Cloud Co-Work: A sandbox web environment with limited access and tools, serving as a better entry point for beginners.

• Cursor: An IDE-integrated tool that has access to open files but operates within certain boundaries.

Installation and Usage Insights

• Installation is straightforward; if users are already subscribed at $20/month, they are encouraged to try out these tools immediately.

Key Concepts in Data Handling

• Important concepts discussed include:

• The context window and data privacy considerations, likened to Dropbox's functionality.

• The iterative process where users describe their needs, cloud code generates and runs the necessary code, followed by user iterations.

Future Directions in Learning

• Upcoming content will focus on practical applications:

• Gathering data from the web to create graphs.

• Demonstrating how powerful these tools can be even without prior knowledge about data sources.