Day 2: Overview

Author
Affiliation

Eduard Bukin

Distributional Impact of Policies. Fiscal Policy and Growth Department

Goals for Day 2

  • Address questions from the self-study period and Day 1.
  • Explore context engineering with #tools, MCP, instructions, /prompts, @agents, skills, and other relevant features.
  • Learn Planning mode to execute complex, multi-step tasks with an AI agent.
  • Delegate a complex analytical task to an AI agent, supervise its execution, and course-correct when it goes off track.
  • Customize your Positron assistant with instructions, prompt files, and skills.

Agenda

Day 2: Advanced AI features

Time Topic
9:00-9:40 1. Welcome and Self-Study Q&A
9:40-10:20 2. Scaffolding on Self-study Example 3 (live)
10:20-10:30 Break
10:30-11:00 3. Tools and MCP (slides + demos)
11:00-11:30 4. Agents and customization (slides + demos)
11:30-12:00 5. Planning mode and exercises (slides + live)

We will use the same examples as in the self-study exercises. Please note that you may need to create several fresh copies of the same example to practice different features:

Before You Arrive

Post your questions, challenges, and discoveries from the self-study exercises in the Teams chat.

Resources

Course data is available in the OneDrive folder: ai4coding-data.

Course examples and materials are available in this repository: ai4coding or on OneDrive: ai4coding-materials.

Day 2: Materials

Slides

View slides in a new tab.

1. Welcome, review of Day 1, and agenda

Review of Day 1 topics, Q&A from self-study exercises, and overview of today’s agenda.

2. Scaffolding on Self-study Example 3 (live)

Live walkthrough of Self-study: Example 3 — reproduce the analysis, check alignment with DIME coding standards, and revise the code using an agentic workflow.

Workflow steps:

  1. Understand — ask the AI to explain inputs, outputs, strengths, and weaknesses.
  2. Adjust and reproduce — change output paths and debug iteratively.
  3. Analyze standards — check alignment with DIME coding standards.
  4. Revise — apply suggested changes (self-study).

3. Tools and MCP (slides + demos)

Explore how tools extend agents with specialized functionality and how MCP (Model Context Protocol) connects AI to external systems.

  • Built-in tools, MCP tools, and extension tools.
  • Using tools explicitly in chat with #tool-name.
  • Permission levels: Default Approvals, Bypass Approvals, and Autopilot (preview).
  • Configuring and using MCP servers (e.g., Stata MCP, Data360 MCP).
  • WB caveat: MCP extensions are not available via the WB Software Center; servers must be configured manually in .vscode/mcp.json.
  • Key distinction: MCP is a protocol exposing tools + resources + prompts; tools are individual callable functions within the IDE.
  • Demo: debugging with tools and MCP side-by-side.

4. Agents and customization (slides + demos)

Understanding agents, subagents, and the agent loop. Customizing AI behavior with instructions, prompt files, custom agents, and skills.

  • The agent loop: Understand → Act → Validate.
  • Agent types: local, background, and cloud.
  • .instructions.md — global rules injected into every request.
  • .prompt.md — reusable task templates invoked with /my-prompt.
  • .agent.md — custom personas with curated tool access.
  • SKILL.md — scoped expertise auto-loaded on description match.
  • Demo: creating instruction files, prompt files, and importing skills.

5. Planning mode and exercises (slides + live)

Use Plan mode to research and design before writing code — a 4-phase workflow: Discovery → Alignment → Design → Refinement.

  • Activating plan mode with /plan in chat.
  • Demo: planning the revision of an analysis do-file.
  • Exercise: plan and execute Example 1 and Example 3 end-to-end.