Today, we’re joined by Anand Naidu, our resident development expert who navigates the worlds of both frontend and backend with ease. We’ll be diving deep into the latest VS Code release, version 1.108, to explore some truly transformative features. Our conversation will cover the groundbreaking new “Agent Skills” that allow developers to teach GitHub Copilot custom tasks, significant refinements to the terminal experience based on user feedback, and streamlined methods for onboarding new team members with shared settings profiles.
The new Agent Skills feature lets developers teach Copilot specialized tasks. Could you walk us through creating a simple skill, detailing the SKILL.md file’s role and how the .github/skills folder structure facilitates this? Please provide a practical example.
Absolutely. It’s a genuinely exciting experimental feature. The process is surprisingly straightforward and elegant. At its core, you create a dedicated folder for your skill inside the .github/skills directory of your project. This folder acts as a self-contained package for the skill’s logic and resources. The real magic, the “brain” of the operation, is the SKILL.md file within that folder. This markdown file isn’t just documentation; it’s a set of instructions that tells Copilot what the skill is, when to use it, and how to execute it. For example, imagine you want a skill to generate a new React component following your team’s specific file structure and boilerplate. You’d create a new-component skill folder. Inside, your SKILL.md would describe the steps: “When the user asks to create a new component, prompt for a name, then copy the template files and rename them accordingly.” The folder would also contain your component templates and maybe a small script, all orchestrated by that central markdown file. It’s a powerful way to embed your project’s domain-specific knowledge directly into the AI.
When a developer makes a request, Copilot now loads relevant Agent Skills on-demand. How does it determine which skill to use from the chat context? Could you elaborate on this detection mechanism and its performance implications for the user experience?
This on-demand loading is a crucial piece of the design, ensuring the system stays nimble. Copilot doesn’t just blindly load every skill you’ve defined. Instead, it actively parses the developer’s chat request and the surrounding context for keywords and intent. It’s looking for triggers that match the descriptions you’ve provided in your SKILL.md files. So, if your request includes phrases like “create a new component” or “scaffold a service,” Copilot’s detection mechanism flags this and checks if any available skills match that intent. If it finds a match, it dynamically injects that skill’s instructions and resources into the current session. The performance benefit here is immense. By only loading what’s relevant for the immediate task, the chat experience remains fluid and responsive. You avoid the overhead of a system bogged down by a massive library of potentially unused skills, making the interaction feel seamless and intelligent rather than clunky and slow.
Terminal IntelliSense defaults were recently reworked based on feedback from power users. What specific friction were they experiencing, and how does requiring a Ctrl+Space trigger better serve their workflow while still remaining discoverable for new users? Please share some insights.
This is a classic case of a feature being incredibly useful for some but disruptive for others. The initial implementation showed IntelliSense suggestions automatically while you were typing in the terminal. For developers who are new to the command line, this is fantastic for discoverability. But for seasoned power users, it was a source of real friction. These are people who live in the terminal; their fingers fly across the keyboard with a muscle memory built over years. An unexpected pop-up, no matter how helpful, breaks that flow and adds visual noise to a space they want to be clean and predictable. The change to an explicit Ctrl+Space trigger is the perfect middle ground. It puts the power user back in complete control—the suggestions appear only when they explicitly ask for them. At the same time, the feature remains discoverable for newer users through improved cues in the status bar, so we haven’t lost the onboarding value. It’s a smart design choice that respects different user workflows.
With the new ability to import a settings profile via drag-and-drop, team onboarding can be streamlined. Beyond just settings, what key configurations would you recommend including in a .code-profile file to quickly get a new team member productive in a specific environment?
The drag-and-drop import for .code-profile files is a massive quality-of-life improvement for team leads. To make a new team member truly productive from day one, I’d go far beyond just cosmetic settings like themes or fonts. The ideal profile should configure the entire development environment. I would absolutely include the list of essential extensions for the project—linters, debuggers, and framework-specific tools. Next, I’d add project-specific code snippets to accelerate common patterns. I would also define the launch.json for debugging configurations and the tasks.json file with pre-configured build, test, and run commands. When a new developer drags that single file into VS Code, their editor is instantly transformed. It’s not just visually aligned with the team; it’s functionally armed and ready to build, test, and debug the project without a single manual configuration step. It turns hours of setup into seconds of work.
What is your forecast for the evolution of customizable AI agents like GitHub Copilot with Agent Skills?
My forecast is that we are on the cusp of moving from generalist AI coding assistants to a new era of specialized, “domain-aware” development partners. Agent Skills is the first major step in that direction. I envision a future where development teams don’t just use a generic AI, but actively cultivate a team-specific AI persona. They will build and share extensive libraries of skills tailored to their company’s proprietary frameworks, internal APIs, and unique deployment pipelines. An agent will know not just how to write code, but how to write code for our specific systems. This will transform AI from a tool that helps with syntax into a true collaborator that understands the company’s architecture and can perform complex, multi-step tasks that are unique to that organization’s workflow. The AI will effectively become an expert on your codebase, radically accelerating both onboarding and day-to-day development.
