Rubber Duck Thursdays!

Key Concepts

• GitHub Copilot CLI: A command-line interface tool for interacting with Copilot, optimized for rapid prototyping and iterative development.
• Contextual Awareness: The principle that AI performance is directly proportional to the amount of relevant information (context) provided by the user.
• AI Infrastructure: The collection of assets—such as instruction files, custom agents, and agent skills—that guide AI behavior within a repository.
• Rubber Ducking: A debugging and problem-solving methodology involving explaining code or logic to an inanimate object (or AI) to clarify thoughts.
• Remote Steering: A feature allowing users to control and monitor local Copilot CLI sessions from remote devices (e.g., mobile phones or web browsers).
• "Make it Work, Make it Pretty": A two-phase development workflow: first, prioritize functionality (MVP/POC), then refactor for readability and maintainability.

---

1. Development Workflow and Methodology

Christopher Harrison emphasizes a "two-phase" approach to coding:

• Phase 1 (Functionality): Use Copilot CLI to rapidly build and iterate. The focus is purely on whether the code works. Harrison notes he often uses the CLI specifically when he "doesn't care about the code" yet, prioritizing speed and real-time feedback (often using Vite for automatic updates).
• Phase 2 (Refinement): Once the feature is functional, move to VS Code to review, refactor, add tests, and ensure the code is clean and maintainable.
• Human-in-the-Loop: Harrison stresses that AI does not replace fundamental engineering practices. Developers must still perform security testing, linting, and unit testing. AI is an "accelerator," not a replacement for developer experience.

2. Managing AI Infrastructure

To ensure AI generates high-quality, project-specific code, Harrison recommends treating AI configurations as first-class repository assets:

• Instruction Files: Markdown files that define coding standards, preferences, and project context.
• Custom Agents & Skills: Specialized tools (found via awesome-copilot.github.com) that can be installed to handle specific tasks like SEO optimization, accessibility (a11y) checks, or Playwright test generation.
• Validation: Use tools like prettier and linters alongside AI to maintain code quality. Copilot Code Review can also be configured to read instruction files to provide feedback aligned with team standards.

3. Real-World Applications and Features

• Remote Steering: Harrison demonstrates using the slash remote on command to control a local CLI session from a mobile device. This allows developers to remain productive while away from their primary workstation.
• Sharing Knowledge: The slash share gist command is highlighted as a critical tool for team collaboration. It allows developers to export the "back-and-forth" history of an AI session, which serves as a training artifact for team members to learn how to prompt effectively.
• Rubber Ducking with AI: By using the "rubber duck" feature, the AI can act as a sounding board, even bringing in secondary models (like GPT) to critique a proposed plan, providing a "second opinion" before implementation.

4. Key Arguments and Perspectives

• Context is King: Harrison argues that most "hallucinations" or poor AI responses are not due to the model's failure, but a lack of context. He compares this to a human conversation: if a request is vague, the response will be off-target. Providing more context (via instruction files or explicit prompts) fixes these issues.
• Don't be Passive-Aggressive with AI: If the AI is capable of finding a file or following a pattern, don't assume it will "just know." Explicitly tell the AI what to do and where to look to guarantee success.
• The Future of Developer Jobs: Harrison posits that the ability to interact with AI—prompt engineering, architectural oversight, and managing AI infrastructure—is becoming a core developer skill. Experience remains vital because it informs the quality of the prompts and the architectural decisions.

5. Notable Quotes

• "I use Copilot CLI when I don't care about the code." (Referring to the initial "make it work" phase of development).
• "AI doesn't change the fundamentals... I'm still going to use tools like prettier. I'm still going to use linters."
• "If you don't know whether or not Copilot's able to do something, ask Copilot."

Synthesis/Conclusion

The main takeaway is that AI tools like GitHub Copilot CLI are most effective when integrated into a disciplined, human-led workflow. By treating AI as an interactive partner that requires clear context and structured "infrastructure" (instruction files and agents), developers can significantly accelerate their output. The transition from "making it work" to "making it pretty" remains the gold standard for professional software development, with AI serving as a powerful force multiplier in both phases.