The Codemap Skill

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.
Up to now Claude has been guessing. It knows we are in VS Code and nothing else. No direction, no goal, no idea why I am here. Time to fix that.
I write a CLAUDE.md. Two parts.
First, the idea. This is my personal fork. I am adding visual elements to help me write high quality code, and I will scope those out later. For now the job is to rip out Copilot and anything that looks or smells like Copilot. Every AI feature goes. I want the file browser, some diff tools, maybe extensions.
I notice there is already an agents.md in there, which is the agent instruction file, and Claude might read it. It is my branch. I get to delete things.
Second, a response style. I drop in a snippet I call desert mode: tokens are water. It is a bit extreme and you can make the obvious joke, that I am spending tokens to tell it not to spend tokens. Fair. But it lasts the whole session, so it pays for itself. I could just write "be terse, bullet points only" and get most of the way there. I want to play with the fun version first.
Quick tangent for anyone new. /clear wipes your context so tokens do not pile up across a session and turn into a bill. I do it constantly. And YOLO is my alias for bypass permissions, which is just do the work. I run in it more than 90% of the time. The exception is plan mode.
Even with a million token window, I do not trust a big context. The bigger it gets, the more hallucinations show up. Keep it small.
Now the map. There is a codemap skill, described as token efficient navigation of large codebases, meant to be used before you explore or grep broadly. Before running it I ask Claude who wrote it, because I have a lot of skills installed globally and I genuinely cannot remember which are mine. It is from Anthropic. Their skills are good, so it gets first shot.
It takes a while and runs a pile of Python to inspect things. Out comes a .codemap directory: MAP.md, recipes, a glossary. The recipes are how you use it, like "add a new workbench feature."
I am not sure I know how to use all of it yet, and I am not sure one map is what I want. So I leave it in place and go build a second one, with a cheaper model, and see which is better.
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.