Q&A: AI-Assisted Data Analysis with Positron

Questions & Answers from February 5, 2026 Seminar

AI
coding
workflow
stata
R
Positron
GitHub Copilot
A compilation of questions and answers from the seminar on using AI agents in Stata and R with Positron. Covers software setup, AI capabilities, confidentiality concerns, and best practices for microeconomists.
Author

Henry Stemmler; Eduard Bukin

Published

February 5, 2026

A. Software

1. How does Positron compare to VS Code, especially with Git?

Answer:

Both integrate well with Git—you get the same source control panel, diff viewer, and commit workflow in either IDE. For Git-heavy workflows, you will not notice any difference.

Key differences:

  • Positron is built specifically for data science, so it includes a native Variables pane, Data Viewer, and Plots panel out of the box

  • R and Python environment configuration is simpler in Positron (automatic interpreter detection, integrated console)

  • VS Code requires more extensions to replicate this functionality and integration is less seamless.

2. Is the Positron set-up similar to AmazonQ for HTML in VS?

Answer:

Not directly comparable. We are not sure if AmazonQ is approved for use at the World Bank, please check with ITS.

3. What are your thoughts on PRISM (OpenAI) compared to the Positron+Copilot set-up?

Answer:

These tools serve fundamentally different purposes. The latter one is not on coding.

4. Does this work with code on Git and data on OneDrive?

Answer: Yes. Git also could be developed on OneDrive.

  • Positron only sees files inside your opened project folder unless you explicitly instruct it to look elsewhere
  • You can keep your Git repository (code, documentation, prompt files) in one location
  • Confidential datasets can remain on OneDrive, a network drive, or another restricted folder outside the project
  • Your code references the data path, but the AI assistant never indexes or “sees” the raw data. AI sees the results of your code, not the underlying files, unless you explicitly share them as context.

5. How do usage limits compare between Positron and VS Code?

Answer:

Usage limits come from GitHub Copilot, not from the IDE itself.

  • If you use the same GitHub account in both Positron and VS Code, you have identical usage limits
  • The IDE is just the interface; GitHub’s servers handle the AI requests
  • WB GitHub is currently on the Copilot Business plan, which offers generous limits suitable for heavy use in data science projects. Use Chat mode for casual use to conserve tokens and Agent when needed.

B. Setup, IT Requirements, Licensing

6. How do you download and install Positron?

Answer:

  • Installation instructions: https://wbggeopov.github.io/geoPovLearn/seminars/seminar-2026-02/setup-instructions.html
  • Currently requires IT installation via eServices until Positron becomes available in Software Center
  • When submitting the eServices request, you will be asked about licensing (see next question)
  • Could be installed by any ITS associate with software installation permissions. Now it is being tested for being included on Software Center for self-service installation.

Steps summary:

  1. Go to eServices and request software installation
  2. Search for “Positron” or submit a custom software request
  3. Follow the prompts regarding licensing
  4. IT will install the application on your machine

7. Do we have a license and what does it cost?

Answer:

  • Cost is $0 — Positron is free, open-source software (built by Posit, the company behind RStudio)
  • When the eServices form asks “Does your unit have a license?”, enter “Yes”
  • This confirms that no paid license is required.

GitHub Copilot is a separate subscription. The World Bank provides Copilot Business licenses. Feel free to reach out to github@worldbank.org

C. Prompting, Personalization, Code Structure

8. Can the AI be personalized so instructions don’t need repeating?

Answer: Yes, through several mechanisms. Read more positron.posit.co/assistant-chat-agents.html

  1. Prompt Files (.prompt.md): Create reusable instruction files that you attach as context when starting a chat. See: https://docs.github.com/en/copilot/tutorials/customization-library/prompt-files

  2. Copilot Instructions File: Create .github/copilot-instructions.md in your repository root. Copilot automatically reads this file and applies the instructions to all interactions within that workspace.

  3. Agent files (agent.md): For complex workflows, maintain a markdown file describing your project structure, coding conventions, and common tasks. Attach it as context when needed.

Example .github/copilot-instructions.md:

9. Does Positron have persistent “project instructions” like ChatGPT?

Answer: Yes, effectively—through the .github/copilot-instructions.md file.

  • Create this file in your repository’s .github/ folder
  • Copilot reads it automatically for every chat and completion in that workspace
  • Instructions persist across sessions without manual attachment

This is functionally equivalent to ChatGPT’s “Custom Instructions” or “Memory” features, but scoped to a specific project rather than your entire account.

Limitations:

  • Instructions only apply within that workspace/repository
  • Very long instruction files may reduce context available for your actual queries
  • Keep instructions concise and focused (recommended: under 500 words)

For personal, cross-project preferences, you can also configure Copilot settings in your GitHub account.

10. When I “always allow” terminal commands, is that permanent?

Answer: It depends on the scope you select:

Scope Duration Risk level
This action only One-time Low
This session Until you close Positron Medium
This workspace Permanent for this project Higher

11. Can we set institutional rules so users do not accidentally allow unsafe actions?

Answer: Yes, and this is strongly recommended for team environments.

Approaches:

  1. Workspace settings (.vscode/settings.json): Pre-configure which commands are allowed or blocked, then commit this file to your shared repository

  2. Documentation and training: Establish team protocols specifying which permissions are acceptable

  3. Prompt protocols: Check the GPID workflow protocols: https://gpid-wb.github.io/Pip-Technical-Guidelines/copilot_protocols.html

  4. Code review: Require review of any AI-generated code that interacts with file systems or external services

D. Confidentiality & Data Security

12. What if I have confidential data on OneDrive?

Answer: The AI only processes files you explicitly provide as context. Several safeguards help protect confidential data:

  1. Project isolation: AI only sees files in your Positron project folder—OneDrive folders outside the project are invisible

  2. Ignore files: Create .copilotignore and/or .aidigestignore in your project root to exclude sensitive files:

    # .copilotignore
data/raw/*
*.dta
*confidential*

  3. Folder structure: Keep confidential data in separate restricted folders, outside your code repository entirely

  4. Explicit context: AI only processes files you attach to a chat or that are open in your editor

  5. Anonymization assistance: You can ask the AI to help write code that anonymizes data before you expose any real values

13. Can AI only see folders we grant access to?

Answer: Yes. The AI has no access outside your Positron project directory unless you specifically direct it there.

E. Stata, R, and Error Handling

14. Can the integrated agent fix Stata errors like it does in R?

Answer: Not currently, due to technical and security constraints: WB imposes policy that forbids running MCP servers locally.

Why R works but Stata does not:

  • Positron has the Model Context Protocol (MCP) server for R that allows Copilot to interact directly with the R console—reading errors, inspecting objects, and running diagnostic code. It also has it for python.

  • Stata lacks an MCP implementation, and World Bank security policies restrict the installation of custom protocol servers.

  • This means Copilot cannot “see” your Stata console output or automatically diagnose errors

Workarounds:

  1. Manual error copying: Copy the Stata error message and paste it into Copilot chat with your code. The AI can still diagnose and suggest fixes.

  2. Log files: Direct Stata output to a log file in your project folder:

    log using "output/analysis_log.log", text replace

    Then attach the log file as context for debugging.

F. Additional Discussion & Comments

15. Comment: “Data management must come first for AI to be effective.”

Response: Agreed. AI amplifies your existing workflow—it cannot compensate for:

  • Missing documentation
  • Unclear data structures
  • Inconsistent variable naming
  • Irregular file organization

Practical implications:

  • Invest time in READMEs, codebooks, and data dictionaries
  • Use consistent naming conventions the AI can learn from
  • Structure projects logically (e.g., data/, code/, output/)
  • Document your workflow before asking AI to help optimize it

Well-organized projects get dramatically better AI assistance.

16. Teams should establish prompt protocols to ensure quality.

Response: Agreed. Standardized prompting improves consistency, reproducibility, and code quality across team members.

Recommended starting point: The GPID workflow protocols provide an excellent foundation: https://gpid-wb.github.io/Pip-Technical-Guidelines/copilot_protocols.html

Key elements of good prompt protocols:

  • Standard prompt templates for common tasks (data cleaning, merging, visualization)
  • Required context to include (data structure, variable definitions, expected output)
  • Review requirements for AI-generated code
  • Documentation standards for AI-assisted work