2026-07-07 | ← Blog
So today was the first time I actually sat down and used GitHub Copilot properly. Actually opened VS Code, looked at it, and tried to understand what I was dealing with.
Honest first impression? Slightly overwhelming. But also kind of fascinating.
I always thought Copilot was just that ghost-text thing that appears when you're typing. You know, the grey suggestion that you hit Tab to accept. I had no idea there were three completely different ways to talk to it.
There's a full sidebar chat (Ctrl+Alt+I) — basically like having ChatGPT docked inside your editor. Then there's inline chat (Ctrl+I) which opens right inside your code file, at your cursor, and the output lands as a diff you can accept or reject. That one genuinely surprised me. And then there's quick chat (Ctrl+Shift+Alt+L), a small floating bar for quick one-off questions that disappears after you're done.
Each one is for a different situation. Sidebar for thinking and planning. Inline for actually writing or fixing code in place. Quick chat for the small doubts you'd otherwise Google.
This one genuinely confused me for a few minutes.
Inside the Copilot chat, there's a model selector dropdown. It lists GPT-5, GPT-4o, Claude Haiku 4.5, Gemini Flash, and a few others. By default it's set to "Auto."
My immediate reaction was — how is Claude here? I never installed Claude. I never set up an Anthropic account. What is happening?
Turns out Copilot is not an AI itself. It's a platform. GitHub has licensing deals with OpenAI, Anthropic, Google — and it routes your request to whichever model is selected. You're always inside Copilot. But the brain answering you could be from any of these companies depending on what's selected.
The "Auto" mode just means Copilot decides on its own which model to use based on your query. Simple syntax lookup? Probably a fast cheap model. Complex reasoning? It'll route to something heavier.
Once I understood this it made sense. It's like an aggregator. The platform is Copilot. The engines underneath are from different vendors.
A few things stuck with me after just one session.
Don't use // comments as prompts inside your code. It works, but it leaves a mess.
Either use Ctrl+I inline chat which doesn't pollute your file, or if you must write
a comment-style hint, use proper Javadoc format (/** ... */) because models are
trained on that pattern and respond better to it.
Slash commands are not just shortcuts — /explain, /fix, /tests, /refactor
are pre-optimized prompt macros. Typing /explain is not the same as typing
"explain this code." The model is primed differently. And combining them with context
like #file:SomeService.java or @workspace makes it actually workspace-aware
instead of guessing.
And the model choice matters depending on what you're doing. For repetitive boilerplate or quick syntax questions, Auto or Haiku is fine. But if you're debugging something weird — a transaction propagation issue, a Spring context loading problem — manually switch to GPT-5 or Claude Sonnet. The difference in reasoning quality is noticeable.
It feels less like an autocomplete tool and more like a second pair of eyes that lives inside your editor. One that doesn't get tired, doesn't judge you for asking obvious questions, and responds in seconds.
I don't fully trust it yet — and I don't think you should on day one. But I can already see how, if used carefully, this changes how much time you spend on the mechanical parts of coding.
I spent some time mapping out all the slash commands, @ agents, and # context variables properly — if you want the full reference I noted it down here: Copilot Commands, Context and How to Use Them Together
More notes coming.