Rubber Duck Thursdays \\ Let's cowork, code, and learn together!

Key Concepts

• GitHub Copilot CLI & SDK: Tools for building agentic workflows and integrating AI features into applications.
• Plan Mode: A CLI feature that allows developers to architect projects and generate markdown-based implementation plans before writing code.
• Rubber Duck Feature: An experimental AI agent (utilizing GPT-5.4) that provides a "second opinion" on development plans to identify potential flaws.
• MCP (Model Context Protocol) Servers: A framework for connecting AI models to local data and tools.
• Awesome Copilot: A community-driven repository for custom instructions, agents, and plugins.
• Local-First Development: A software design philosophy prioritizing local data storage and offline capabilities, often using JSON or SQLite.

1. GitHub Platform Updates (April Changelog)

The speaker highlighted several new features released to the GitHub platform:

• Comment Management: Users can now hide comments with specific reasons (e.g., spam, off-topic, duplicate, resolved).
• Notification Sorting: A highly requested feature allowing users to sort notifications by newest/oldest.
• Contributor Roles: Pull requests now display labels identifying the contributor's role (e.g., regular contributor, repository member, or bot like Dependabot).
• Issue Sidebar Enhancements: Improved visibility of release tags and default values directly within the issue sidebar.

2. GitHub Copilot SDK & CLI Workflows

The speaker demonstrated how the SDK extends the CLI to build custom AI-powered tools:

• Agentic Capabilities: Developers can create "agent sessions" to automate tasks, such as building features, updating dependencies, or generating documentation.
• Real-World Applications:
  • Browser Automation: An agent that navigates a shopping site, adds items to a cart, and sends a notification.
  • Workflow Automation: A tool that generates PowerPoint presentations from CLI session logs and schedules stakeholder meetings.
• Model Flexibility: The platform allows users to switch between different models (Claude Haiku, GPT-4, etc.) within the same session to leverage their specific strengths (e.g., Haiku for fast planning, GPT for backend logic).

3. Step-by-Step: Using "Plan Mode"

The speaker outlined a methodology for using the CLI as a "thinking tool" rather than just a code generator:

1. Initiate Plan Mode: Use Shift + Tab in the CLI to enter the planning environment.
2. Define Requirements: Input the project context (e.g., "My app uses Electron and React").
3. Iterative Socratic Dialogue: The agent asks clarifying questions (e.g., "Do you prefer JSON or SQLite for storage?") to help the developer refine their architecture.
4. Review & Refine: The agent generates a markdown-based plan. The developer can manually edit this document or ask the agent to adjust specific sections.
5. Execution: Once the plan is finalized, it can be handed off to specific agents (e.g., an accessibility expert or a backend specialist) to implement the code.

4. Customization and Community Resources

• Awesome Copilot Repository: A central hub for community-contributed custom instructions and agents.
• Custom Agents: These are essentially "glorified markdown files" containing front matter (name, description, model preference) and specific system instructions.
• Example: An "Accessibility Expert" agent that enforces specific web standards by scanning code against defined accessibility rules.

5. Notable Quotes

• "I honestly use plan mode more than I actually write code with the CLI just because using it as a thinking tool is a very helpful exercise."
• "I do think personalized solution is the wave of the future... everything you build can be very custom to you."
• "Our evaluations show that Claude Sonnet and the rubber duck [agent] makes up 74.7% of the performance gap between Sonnet and Opus alone."

6. Synthesis and Conclusion

The session emphasized a shift toward "Personal Software," where developers use AI agents to maintain and extend older projects (like the speaker's 10-year-old "To-Do Meter" app) that they previously avoided touching. The core takeaway is that the combination of the Copilot CLI, SDK, and Plan Mode transforms the development process from manual coding to architectural orchestration. By using AI as a Socratic partner to vet plans and handle maintenance chores (like dependency updates), developers can maintain high-quality, custom-built applications with significantly less overhead.