Build with the Copilot CLI - Mona Mayhem

Key Concepts

• Agentic Development: Using AI agents to perform complex, multi-step coding tasks autonomously.
• Context Engineering: Providing AI with specific instructions, references, and constraints to improve output quality and reduce hallucinations.
• Plan Mode: A workflow where the AI generates an engineering document/plan before implementation, allowing for human review and refinement.
• Autopilot: An autonomous mode where the AI iterates through tasks, runs tests, and verifies results without constant manual input.
• YOLO Mode: A mode that automatically approves tool calls (like terminal commands) to speed up development.
• Fleet Mode: Running multiple agents concurrently to handle different tasks in parallel.
• Delegate: Offloading tasks to a cloud-based agent that works on GitHub issues independently, resulting in ready-to-review Pull Requests (PRs).
• Human-in-the-Middle (HITM): The essential practice of human oversight to ensure AI outputs meet quality, safety, and design standards.

---

1. Main Topics and Workflow

The workshop focuses on building "Mona Mayhem," a retro-themed arcade game that compares GitHub contribution data between two users. The development process emphasizes:

• Initialization: Using copilot init to scan the repository and generate a copilot-instructions.md file, which acts as a "grounding document" for the AI.
• Planning: Using plan mode to define route structures, error handling, and caching strategies before writing code.
• Implementation: Utilizing autopilot to execute the plan, followed by iterative refinement.
• Refinement: Using screenshots and direct feedback to fix UI/UX issues (e.g., responsiveness, animations).

2. Methodologies and Frameworks

• Contextual Documentation: Beyond general instructions, developers can create Skills—dynamic, task-specific bundles (e.g., accessibility checks or unit testing) that the AI can pull from as needed.
• Token Optimization: To manage costs and efficiency, developers are encouraged to be highly specific in prompts and use lighter-weight models (like Haiku) for simple tasks, reserving more powerful models for complex debugging.
• Orchestration: When using Fleet mode, the AI manages potential race conditions by "pressing the brakes" on agents if they attempt to access the same file simultaneously.

3. Key Arguments and Perspectives

• Non-Determinism: The presenters emphasize that AI models are non-deterministic; the same prompt may yield different results. This makes "Human-in-the-Middle" oversight critical.
• Responsibility: Ayan and Matt stress that developers remain responsible for the code produced by AI. If a manager reviews poor code, the AI cannot be blamed; the developer must exercise "taste" and judgment.
• Iterative Design: The presenters argue that the most effective way to use AI is to treat it as a collaborator, using "banter" and iterative feedback to polish the final product.

4. Notable Quotes

• "The goal is to create artifacts, create instructions that AI is able to reference so it doesn't have to reinvent the wheel over and over again." — Ayan
• "If I'm your manager... and you come out with something that's iffy, you can't blame the AI for it. You took responsibility for it." — Matt
• "The ability to use AI to develop thoughtfully designed plans... is one of the greatest skills that any developer can have in today's day and age." — Ayan

5. Technical Tools and Commands

• copilot init: Scans the repo to create a tailored instruction set.
• ! [command]: Allows interaction with the shell directly from the Copilot interface.
• copilot plugin marketplace list: Used to discover and add community-based bundles like "Awesome Copilot."
• shift + tab: Toggles between different AI modes (Plan, Autopilot, etc.).

6. Synthesis and Conclusion

The workshop demonstrates that modern agentic development is less about writing code and more about orchestration and design thinking. By utilizing planning phases, maintaining clear instruction files, and employing multiple agents (Fleet mode) for parallel tasks, developers can significantly accelerate their workflow. However, the core takeaway is that AI is a tool that requires human guidance, specific constraints, and constant verification to ensure the final output is functional, accessible, and aesthetically sound.