Workshop
Agentic Coding in VS Code
VS Code Copilot's agent mode has a deep customization surface — instructions, prompts, skills, agents, subagents, hooks, MCP — but the docs are scattered and the concepts don't obviously connect. This workshop builds the mental model from scratch, walks through two real projects, and ends with hands-on group work. No vibe-coding: every concept earns its place.
Companion reference: The full Agentic Coding Best Practices doc is embedded at the end of this workshop — a comprehensive cross-platform reference covering Claude Code and VS Code Copilot together, including anti-patterns, model selection, and the complete project pipeline.
Who this is for
This workshop is designed for three types of people who keep running into the same wall with VS Code Copilot’s agent mode:
- Software engineers who use VS Code Copilot for inline suggestions or basic chat but have never customized an agent, written a skill, or set up a hook
- PMs and non-coders who want to run structured agentic workflows without writing code from scratch — and who keep hearing “just vibe-code it” as the answer
- Anyone who has read the VS Code agent documentation and found the concepts hard to connect into a working mental model
What you’ll have by the end
Three concrete deliverables — one per layer of the reliability stack:
- A working
.github/copilot-instructions.mdthat constrains the agent’s behavior for a real project context — the foundation every other asset builds on - A reusable
.prompt.mdfile that captures a repeatable workflow as a named/command— the first step up the reusability ladder - A
SKILL.mdbundled with a script and a template — a multi-step capability that loads on demand and can be moved across projects
Pages
01
Setup
Account requirements, extensions, agent mode toggle, MCP scaffolding. Five checks before the first prompt.
02
Concepts
The full customization surface — all 9 file types explained: where they live, what they do, when to use each. Mapped to the 5-layer reliability stack.
03
Connecting the Dots
Decision tree, reusability ladder, lifecycle pipeline, and 10 anti-patterns with fixes.
04
Project: Sprint Planning
Build a sprint-planning agentic workflow end-to-end — no code required. Instructions → prompt → skill → agent → hook.
05
Project: EDA Analyser
Build a Python EDA CLI with guardrails — instructions, skills, a critic subagent, DuckDB MCP, and a formatting hook.
Ref
Best Practices Reference
Full companion document. Cross-platform framework, anti-patterns, model selection, quick-reference cheatsheet.