Auditing the Workflows

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.
Skills, instructions, prompts and commands are out of .github and into .claude. I am committed to running this whole project through Claude Code. What is left is the issue template and the GitHub Actions.
Before I touch it, /usage.
Four dollars for that entire last lift. 8% of the session. You hear horror stories about people and their giant codebases. I feel fine. I expect to run out of energy before I run out of session.
You might be wondering if I am winging this. I am not. I have thought about it a lot and I have stopped and started a few times. But I also do not think any of us really know what we are doing with a codebase like VS Code, which is enterprise and open source at the same time. So I am running two competing instincts: understand what is there before changing it, and do not fall into analysis paralysis. Branches are what let me hold both. If I screw up, that is fine. That is how this works.
One more thing first: /clear. I am not letting this context pile up.
Then the ask. I have cleared a lot out of .github, but the GitHub Actions are still there and I do not know what they do or whether I need them. I suspect they are for public PRs and team process. Audit them. Tell me what is there and what I can get rid of. I do want a test surface and linting, so prune down to only what I need.
The audit:
- Most are pinned to self-hosted runners. On my fork they would queue forever and never run. Dead on arrival.
- One actually works, but it ships Monaco, and I do not ship Monaco.
- What to keep? Nothing.
Feels good. Nuke it and replace it with one ci.yaml that runs the surface I asked for: ESLint, styling, hygiene, type check client, validate layers, test node.
Thirty-one files. Everything passes, plus a few latent breakages found and fixed along the way. And one detail I loved: the post-install step symlinks the Copilot instructions. So my suspicion was right. The VS Code team is running Claude Code against their own Copilot instruction files.
One workflow file. No issue template, because I am not Microsoft. Ready to go.
A quick aside, since it came up. I worked with a developer once who was relentlessly polite to Claude. Everyone gave him grief about wasting tokens on "please" and "thank you." He said, no, in the robot uprising they are going to remember this. So you never know. I have also found that complimenting Claude gets you slightly better work. Maybe that is me. Either way, spreading a little joy is what life is about.
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.