Mining Their Instructions

Drive AI agents through a million lines of code you did not write. Recon, surgical navigation, codemods, characterization tests, and a toolkit of four reusable skills you keep.
I just cleared every Copilot artifact out of .github. Prompts, commands, instructions, skills, all of it.
And then it hit me. The VS Code team uses Copilot to work on VS Code. Which means those files are the team writing down, over years, how their own codebase is supposed to work. Architecture rules. Layering rules. How the tree widgets work. Design philosophy.
That is not Copilot garbage. That is the manual, and it works just as well in Claude Code.
So: go into main, read .github, and give me a rundown of every Copilot artifact in there. What is useful for my work, and what is only there for the Microsoft team? Output it on screen, do not write any documents.
A note on that prompt while it runs. Normally I use a three-step structure: you are a, I need a, give me this. I have been softening on that as the models get smarter. The context is already in CLAUDE.md, so it knows the project and the goal. And because I am using WhisperFlow I can pack a lot of intent into two spoken paragraphs.
You might be thinking that talking like that burns tokens. It does. Two paragraphs against a codebase this size is nothing, and the extra context makes everything downstream terser. Good trade.
The rundown comes back sorted:
- Generic craft, useful to you. Coding guidelines, best practices, source code organization, disposable observables, design tokens, design philosophy.
- Area specific, only if you touch it. You said you want the file browser, so here is how the trees work.
- Skip. Notebooks. Sessions. All the team process material.
- Grab. Design, UX theming, UX layout, and the prompts.
I take the recommendations, because it already knows tokens are top of mind and it is filtering with that in mind. Extract what is worth keeping into a .claude directory so I can reuse the commands and skills, and scrub the parts that constrain me to their process. I am not a corporation. But I do want the output to be good enough that someone contributing knows what they are looking at.
Out comes a set of skills: architecture, visual, test guidance. Some will stay, some will go. Every Copilot agent reference stripped.
One small thing it flags: an uncommitted deletion that predates this session, left over from the strip. That was me.
Commit it and move on. One thing left in .github: the workflows. I do not think I need them. But I need to know what they do before I say that.
Chesterton's fences.
Drive AI agents through a million lines of code you did not write. Recon, surgical navigation, codemods, characterization tests, and a toolkit of four reusable skills you keep.