851: The Future of VS Code and Copilot

Episode Summary

Executive Summary: Recorded at GitHub Universe, this episode focused on the latest VS Code and Copilot announcements, especially how AI features are becoming more practical, controllable, and editor-native. Cassidy Williams and Harold Kirschner discussed live demo reliability, multi-file editing, custom instructions, inline chat, experimental AI search, repo indexing, and Copilot Extensions, emphasizing that UX and context delivery matter as much as the model itself.

Main Topics: GitHub Universe keynote and live demo reliability: Cassidy described the pressure of a fully live AI demo and the extensive backup plan used to ensure the keynote worked despite AI nondeterminism. Copilot custom instructions and prompt shaping: The guests explained how markdown-based repo instructions let developers steer Copilot toward code style, dependencies, conventions, and project-specific behavior. Multi-file editing and new editing workflows: They highlighted multi-file edits as a major feature, enabling creation and modification across files with diff-based review and selective acceptance. Inline chat, chat UX, and AI interaction patterns: A major theme was the importance of choosing the right UI for the task—inline chat, sidebar chat, terminal chat, and workspace flows all serve different levels of confidence and complexity. Experimental AI features: search, review, repo indexing, next-edit suggestions: The discussion covered upcoming and experimental capabilities such as semantic search, code review assistance, next edit suggestions, and temporal/repo context awareness. Copilot Extensions and bringing AI into external systems: They explained how extensions let teams connect Copilot to custom tools, APIs, issue trackers, and enterprise workflows, extending Copilot beyond generic code assistance. Cross-editor expansion and future platform strategy: The conversation ended with Copilot expanding to Xcode and other ecosystems, reflecting a push to make AI assistance available wherever developers work.

Key Arguments: Live AI demos require multiple layers of fallback because model outputs are non-deterministic and can vary from rehearsal to rehearsal. Custom instructions are a practical way to encode repo-specific conventions, dependencies, and corrections without overloading the model with irrelevant context. The value of AI coding tools depends heavily on UX: inline chat, multi-file diffs, and explicit context selection can make AI assistance feel understandable and safe. Semantic or repo-aware context is more important than raw model capability; the tool needs the right data and interaction flow, not just a better model. Multi-file editing is especially useful for refactoring tasks that span schema, backend, UI, and state management, where simple rename tools are insufficient. Copilot Extensions allow organizations to connect AI to internal APIs and systems, making Copilot useful for specialized enterprise workflows. The team is actively experimenting because the best AI editor patterns are still evolving, and user feedback plus research drive what ships.

Data Points: Backup layers in keynote demo: 4 - Cassidy described the demo as having a video fallback, a second laptop, the live presenter, and "a hope and a prayer." Copilot Extensions release timing: May - Harold said Copilot Extensions was released in May and is still in public preview. Talk/demo planning window: Days or weeks - Harold noted the demo changed repeatedly in the days or weeks leading up to the keynote because features and UI were still moving fast. User feedback cadence: Daily - They mentioned VS Code Insiders and Copilot Chat updating every day in some cases to stay in sync with rapid changes. Context bubble for chat: Present in current experimental settings - They described a new chat context bubble that lets users control what goes into chat.

Pivotal Quotes: "We had the video, we had the second laptop, we had me, and then a hope and a prayer." — Cassidy Williams: Describing the layered backup plan for the live AI keynote demo. "The biggest, coolest one is probably multi-file editing." — Cassidy Williams: Summarizing the headline Copilot feature announced in the episode. "The models are surprisingly almost table stakes now; it's how you give the data to the LLM and how you then apply it." — Host/Discussion: Framing the conversation around UX, context, and product design rather than model quality alone.

Implications: AI coding tools are shifting from novelty to workflow infrastructure. The winners will be the tools that combine strong models with great context, clear diffs, and low-friction UX across editors, terminals, and enterprise systems.