Revolutionize Your Analysis in Stata and R

AI Agent-Assisted Workflow with GitHub Copilot and Claude

Eduard Bukin ebukin@worldbank.org
Distributional Impact of Policies
Fiscal Policy and Growth Department

2026-02-05

Motivation

Chat and Web-Based AI tools are impressive!

ChatGPT | Copilot | Gemini | WB MAI

Use familiar technologies:

• Web-Browser,
• Chat,
• Stata editor,
• copy-paste …

Is this the best way to use AI for data analysis?

In fact, there are many
AI-powered
Integrated
Development
Environments (IDEs)
for coding and data science!

The goal

of this seminar is to introduce you to AI-assisted data analysis with Positron IDE and GitHub Copilot | Claude.

There are many IDEs ➔

Agenda

• Introduce several AI-Concepts (vocabulary)

• Share experience of using AI-assisted workflow in Positron with Stata and R

• Provide kick-off instructions and resources.

Key Concepts

What do we need to know about modern analysis with AI?

1. AI Integrated IDE: Chat | Agent | Inline Completion
2. Context awareness: How AI understands your project
3. Model Context Protocol (MCP): Universal adapter for AI
4. GitHub Copilot | Claude: LLM providers
5. Efficient prompting: Getting the best results
6. Caveats and limitations: What to watch out for

AI Integrated IDE

AI: Chat

Positron Assistant

• Ask AI (Claude 4.5) through Github Copilot

• Provides explanations, suggestions, and code snippets.

• Integrates with project context, and code.

• Learn more:

  Assistant Chat

AI: Agent

Positron Assistant

• Executes instructions.

• Acts independently

  • Runs code
  • Fixes errors
  • Learns
  • Reasons

• See more in the live demo!

AI: Inline Completion

Positron Inline Code Completion: Suggests code snippets as you type.

Context Awareness

Positron accesses project metadata. Thus AI ‘knows’:

• Files str.: Code, docs
• Data: Var. names, types
• History: Edits, commands
• Environment: Packages
• Intent: Current task
• Results: Output, errors

Why does it matter?

• Project-specific suggestions
• Understands dependencies
• Reduces hallucinations
• Improves efficiency

Model Context Protocol (MCP)

MCP is a universal adapter for AI—Anthropic— that connects data flows:

• Stata-MCP or R MCP for Stata/R ⟶ Positron ⟶ Copilot ⟶ LLMs.

GitHub Copilot | Anthropic Claude

GitHub Copilot

• WB-approved
• github.com/worldbank
• Uses Claude | GPT
• Integrates with IDEs
• github@worldbank.org
• AI @ WB: ai.worldbank.org

Choose your LLM:

• Claude Sonnet/Haiku/Opus
• OpenAI GPT-4/o1…

Efficient Prompting

• Be specific:

  “Write a Stata do-file to …” / “Refactor this R function to …”

• Provide context:

  “Goal: X; Dataset: Y variables; Constraints: Z (WB rules, packages, runtime)”

• Define expected output:

  “Save as regression_results.xlsx, format as APA table” / “Create bar chart with 95% CIs”

• Summarize + clarify first:

  “Restate and ask clarifying questions before implementing”, “Explain why …”, “Give alternatives with trade-offs…”

• Iterate in small steps:

  “minimal changes”, “refine”

• Set boundaries:

  “Don’t use … data”, “Don’t print secrets, ask if in doubt.”, “Don’t change files”.

Limitations and Remedies

• Wrong-but-plausible outputs / hallucinations: code runs but logic is wrong

  Verify and validate: ask the model to explain and justify the solution

• Context limits: not all files/data are in context; too large projects.

  Be explicit: state assumptions, expected inputs/outputs, and references

• Outdated knowledge: suggested APIs/packages/options may have changed

  Teach the model: provide references/links; ask it to learn

• Over-reliance: erodes fundamentals; mistakes slip through unchallenged

  Keep learning: ask for step-by-step reasoning; request alternatives and trade-offs

• Confidentiality / security / privacy

  Constrain context: exclude sensitive data; use .copilot-ignore; AI @ WB

• Reproducibility: answers can vary across sessions/models/settings

  Cutomize agents: save prompts, use Git; create AI agents

Summary

Why use IDEs, not a web-browser-based workflow?

• Context-awareness
• Streamlined workflow
• Reduced friction

Why Positron?

• Built for data science, not software development
• Integrates with Stata, R, and Python seamlessly
• Advanced AI features for data analysis

Where to Start?

• Setup the software: follow instructions

• Reproduce demos:

  • Live demo in Positron with Stata (recording) + Materials

  • Posit Conf 2025 Postron assistant Demo in R and Python

• Ask AI’s help to learn.

• Step out of your comfort zone:

  • learn and experiment with new technologies: Git, GitHub, R, Python—they are there for a good reason!

• See more slides with additional materials below.

Live Demo

From an old analysis in Stata to an upgraded Stata+R reproducibility package in under 10 minutes!

• Download materials here: github.com/WBGGeoPov/seminar-coding-with-ai-demo

• Watch the seminar recording: Positron with Stata (recording)

• Seminar internal page: link

Tip

Ask AI for help: “How do I download a project from GitHub and open it in Positron IDE? The link is: …”

Live Demo: Positron IDE overview

Thank You! Questins?

Resources & Source Code

• Internal seminar page

• Recording

• Positron + Stata: Live Demo (code)

• Slides and Supporting Materials

  • Slides (web) + Slides (PDF)

  • Setup Instructions

• Presentation Source Code

• GeoPov learning resources

Contact: Distributional Impact of Policies poverty@worldbank.org

Additional materials

Software Setup Overview

Note

Full details: Setup Instructions

1. Install prerequisite software (via WB Software Center)
   • Stata 19+, R 4.5+, Python 3.13+, Quarto, Git
   • Install Python uv package: pip install uv
2. Install Positron IDE (system-level install): Request help from IT if needed
3. Install key extensions in Positron
   • Stata MCP, Quarto
4. Connect GitHub and configure Positron Assistant
5. Start experimenting!
   • Open assistant: Ctrl+Shift+P > “Ask Positron Assistant”
   • Try: chat, agent mode, inline code completion

Positron IDE: Self-learning

Positron: Modern AI-native IDE for data science

• Built by Posit (creators of RStudio)

• Supports Stata, R, Python, and others

• Watch the introduction video →

• Try it yourself with examples in R and Python:

  • github.com/posit-dev/posit-conf-2025-positron-assistant-demo

Positron: Assistant

Positron: Data explorer

Positron + Stata

1. Make sure prerequisite software is installed (Stata, R, Positron)

2. Install Python after that the uv package: pip install uv

3. Install Stata MCP in Positron and configure Stata path and Edition

4. Create a new Stata do-file, write some code, save it and press run it.