SUMM

The talk focuses on "VIP" coding at scale, prioritizing output over code details.
Emphasizes embracing exponential advances, building trust in AI, and adding guardrails rather than reviewing every code line.
Discusses the evolution of coding workflows with AI, referencing trends like increasingly autonomous agents generating code.

Outlines three stages: "Yolo Vibes" (rapid prototyping, creativity), "Structured Vibes" (balance, maintainability), and "Spectrum Vibes" (best practices for scale and reliability).
"Yolo Vibe" coding emphasizes speed and creativity, suitable for prototyping, learning, and personal projects.
Larger companies use "vibe" coding to help non-technical people communicate ideas and quickly create mockups.
"Structured Vibes" focus on code maintainability, readability, and handover quality.
"Spectrum Vibes" incorporate emerging best practices for scaling with reliability and speed.

Demonstrates starting YOLO VIP coding with GitHub Copilot and VS Code, using a new workspace in agent mode.
Suggests disabling the scaffold workspace tool for purer HTML prototyping, but notes current challenges in disabling it.
Emphasizes choosing popular, stable frontend stacks (e.g., React with Vite, Material or Fluent design) to maximize AI's effectiveness.
Shows how to prompt AI for attractive, accessible UI, leveraging design principles (Apple, Material Design).
Guides through setting "auto approve" to let AI proceed without user intervention, increasing coding speed and flow.
Illustrates that AI can generate a hydration tracking app with minimal user input, focusing on high-level requirements.

Explains workspace vs. user-level auto-approve settings.
Shows handling multiple coding sessions with differing settings, and the benefits of letting AI try various approaches.
Highlights Copilot’s ability to interpret minimal design instructions and generate features (e.g., metrics in metric units, visually appealing elements).
Discusses the "work visually" feature: selecting an element in browser preview and prompting changes (like adding particle effects) without needing to reference code directly.
Demonstrates easy undo through checkpoints for reverting AI-generated changes.

Describes features supporting YOLO coding: agent mode, flexible panel/chat layout, multi-window and chat management.
Introduces voice dictation for prompt inputs, supporting quick and accessible interaction with Copilot.
Notes the value of high-level prompting to test AI’s design strength and using known technologies for more reliable outputs.
Recommends using the workflow for ideation, experimentation, and not becoming too attached to early results.

"Structured VIP" stage balances creativity with reproducibility, embedding guardrails for consistency and maintainability.
Companies leverage starter templates with clear instructions and constraints to align AI activity with organizational standards.
Workspace instructions can enforce specific stacks, design systems, and best practices, ensuring consistent outcomes across teams.

Shows using Copilot to write commit messages and push code to GitHub.
Discusses using GitHub Codespaces and devcontainers for reproducible development environments.
Stresses the value of a "copilot-instructions.mmd" markdown file for context-aware AI guidance (stack info, conventions, tool preferences).
Mentions new dynamic instructions: instructions can be scoped to file types, with glob patterns, though some limitations exist.
Introduces workspace/user-level and reusable team prompts for tasks like writing tests.
Demonstrates creating and sharing prompts for consistent team workflows.

Describes new ability (in Insiders) to create custom chat modes (e.g., a TDD mode) that influence AI behavior for the session/project.
Custom modes can be configured at user or project level and fine-tuned for tasks and tool use.
Demonstrates building a TDD mode: prompt AI to generate failing tests first, confirm completion, then implement features.
Highlights need for explicit tool configuration within prompts and plans for higher quality/consistency.

Explains adding, configuring, and managing MCP (Multi-Command Protocol) servers to augment the AI with custom tools for research, browser automation, database access, etc.
Shows tool configuration via JSON or install dialogs, including secure handling of API tokens with encrypted storage.
Discusses server types, protocols (SSE, HTTP), and caching of available tools.
Notes the need to be in agent mode (not ask mode) for full tool utilization.
Users can manage multiple AI models (e.g., local, third-party, custom-API) through a model picker.

Recommends minimizing tool clutter for deterministic results in custom modes (e.g., restrict a research prompt to only use perplexity engine).
Users can add specific tools to context for precise control.
AI may not always use suggested tools; tool-calling is ultimately AI-mediated.

Outlines mature workflow: start with workspace and dynamic instructions, use custom tools and prompts, leverage planning/spec modes.
Shows integrating external resources (web docs, cross-repo search) for richer AI responses.
Stresses frequent commits, iterative feedback, and using pause/checkpoint features to maintain process control.
Encourages experimenting to find the best personal/team process involving modes, prompts, and instructions.
Highlights AI's potential as a critical design/thought partner (not just code-generation), prompting it to critique, question, and help plan projects.

Final advice: experiment, iterate, and find your optimal workflow.
Use structured, well-documented, and self-explaining code bases with clear, updated instructions for best AI performance.
Keep evolving team practices with custom modes, prompts, and clearly defined instructions to suit changing needs.

Real World Development with GitHub Copilot and VS Code — Harald Kirschner, Christopher Harrison