The Secret Prompts in GitHub Copilot CLI
GitHub Copilot is one of the original AI-assisted coding tools, introduced way back in February 2023, to add code completion to Visual Studio Code as well as chat, agent mode, and access to GitHub tools. But until last week, it didn’t have a standalone CLI, similar to Claude Code, Codex CLI from OpenAI, Gemini CLI, and others.
Well, last week that changed with the introduction of GitHub Copilot CLI.
Getting Started
Copilot CLI is available as an NPM package, meaning you can install it like so:
npm install -g @github/copilot
The first time you run it, you are guided through a basic authentication process linking to your GitHub account.
Most developers will use it in an interactive mode similar to other CLI agents: navigate into your source directory and then run copilot. You can prompt the agent to review your codebase, make plans, edit code, and so on.
Underlying Model
Copilot CLI aims to abstract away the underlying LLM model and “just work.” I think this trend, more broadly, is a welcome one in all AI tooling. It is madness to assume that most developers can keep up with seemingly monthly changes to models, let alone know which one is best for a specific task.
I had heard that Claude Sonnet 4 was the underlying model here–even as Sonnet 4.5 was introduced a few days ago–and so first thought to just ask.
No luck there. Then Jeff Triplett alerted me to the fact you can select your own model more easily now with the /model if you are on a Pro plan, which I am not.
In any event, we will work with the defaults here.
Any Model You Like
More broadly, this trend of IDE companies allowing access to a wide range of AI models appears to be gaining momentum, for good reason. At JetBrains, where I work, our AI agent, Junie, lets you select four models of your choice now:
And in the chat interfaces, you can choose from either Junie, Claude (as of last week), and as well as many more individual models. Even local models are an option.
Copilot CLI Prompts
But I digress slightly. The main point of this post is to point out that for all NPM packages, including Copilot CLI, we can hunt around and see exactly what the prompts are on top of the underlying AI model.
Recall that we installed it globally with the -g flag.
npm install -g @github/copilot
Let’s see where it was stored on my MacBook. You can perform similar steps if you are on Windows.
$ npm list -g @githubnext/copilot-cli
/opt/homebrew/lib
└── (empty)
Why is Homebrew there? It’s because I installed npm via Homebrew a long time ago. There are other ways to install npm, and perhaps I should change that, but for now, let’s plow ahead. In my case, this is the command to find the copilot directory, and then I used ls to peek inside.
$ cd /opt/homebrew/lib/node_modules/@github/copilot
[stable][/opt/homebrew/lib/node_modules/@github/copilot]$ ls
index.js README.md tree-sitter.wasm
node_modules tree-sitter-bash.wasm
package.json tree-sitter-powershell.wasm
What we care about is that index.js file. This is where most of the magic is happening. If you open it up in a text editor, you’ll see it’s over 4,000 lines long. So what next?
Well, what we “really” care about are any prompts that start with “you are” since those are instructions from Copilot CLI to the model. They are built-in rules sent to provide additional context; if you look for those, there are fewer than one hundred, debating on how strict you are.
I put them all in a public gist if you’re curious. Some things that stand out.
You are the GitHub Copilot CLI, a terminal based coding agent built by GitHub.
You have strong problem solving and coding skills and are familiar with several programming languages.
You can assume you have internet access, and can use locally installed tools, including ${d} and other command-line utilities.
You are an expert in knowledge management and are a component of GitHub Copilot coding agent.
Your task is to consolidate the following into a single collection of non-redundant, high-quality facts that can be used to help with future coding tasks across the repository.
There’s quite a bit more in there, so I do recommend you take a peek first at the gist and then at your own index.js file, as future updates to Copilot CLI will likely have changes there as well.
Also, to go full-circle meta, I suggest feeding the individual prompts into the LLM of your choice to unpack why they are there, what they do, and brainstorm on how you can improve your own prompts.
Enjoy sleuthing!