Navigate a large codebase with Claude Code

Where you launch Claude decides which CLAUDE.md files load and which files it can edit.

1. 1In a terminal, cd into the package or subsystem you work on, such as packages/api.
2. 2Run claude, then type /init at its prompt.
3. 3Read the CLAUDE.md it writes, and delete anything Claude could work out from the code.

You should see a new CLAUDE.md in that folder, with its build and test commands.

Started from a subfolder, Claude can edit only that part of the repo until you grant it more.

The root CLAUDE.md loads in every session, so each extra line costs attention on every task.

1. 1Open CLAUDE.md at the repository root.
2. 2Keep rules that hold in every folder: commit style, generated files and things never to touch.
3. 3Delete folder maps and architecture notes. Claude reads the repo itself, and Anthropic's docs now list that habit as old advice.

# Rules for every folder
- Run scripts from the folder you changed, not the repo root.
- Start commit subjects with the folder name, like `api: fix login timeout`.
- Never edit files in [GENERATED FOLDER]. Run [CODEGEN COMMAND] instead.

You should see a root file of a few lines, far under the 200 lines Anthropic suggests as a limit.

A folder's CLAUDE.md loads only when Claude works there, so it runs that folder's tests instead of the whole suite.

1. 1Add a CLAUDE.md to each busy folder, such as packages/web/CLAUDE.md.
2. 2List how to test one file, run the folder's tests and lint, as in the example.
3. 3Start a session in that folder, run /context all and check the list under Memory files.

# packages/api
- Test one file: npm test -- [PATH TO TEST FILE]
- All API tests: npm test
- Lint: npm run lint
- Copy .env.example to .env before you run tests.

You should see the root CLAUDE.md and your folder's file under Memory files, and none from other folders.

A language server lets Claude find a function's definition and callers by symbol, instead of reading file after file.

1. 1Install your language's server first. For TypeScript, run the first line below in your terminal.
2. 2Start claude and run the /plugin install line at its prompt.
3. 3Other languages have their own plugin, such as pyright-lsp for Python or gopls-lsp for Go.

npm install -g typescript-language-server typescript
/plugin install typescript-lsp@claude-plugins-official

You should see a message that the plugin is active, or that you need to run /reload-plugins first.

Code intelligence works in terminal sessions. Cloud sessions don't start plugin language servers.

Deny rules stop Claude from opening big files that were never meant to be read.

1. 1Open .claude/settings.json in the folder you start Claude from, or create it.
2. 2Add the rules below, and change the paths to match your build folders.
3. 3Commit the file so your whole team gets the same rules.

{
  "permissions": {
    "deny": [
      "Read(./**/dist/**/*)",
      "Read(./**/build/**/*)",
      "Read(./**/*.generated.*)",
      "Read(./**/vendor/**/*)"
    ]
  }
}

You should see your 4 new rules listed in /permissions, with the settings file each one came from.

Paths in .gitignore, like node_modules/, already stay out of Claude's searches.

One trace across folders shows whether the setup works before you rely on it.

1. 1Paste the prompt below, with a real function name from your code.
2. 2Watch which files Claude opens. Each one should sit on the path you asked about.

Trace how [FUNCTION NAME] gets called, from the entry point down to where it does its work. Use the language server to find references. List each file and line in order. Don't edit anything.

You should see a short, ordered list of files and lines, with no detours through build output.