vagejs with Claude Code
Anthropic's agent in your terminal.

Set up

Install Claude Code, create a site, and start a session inside it:

curl -fsSL https://claude.ai/install.sh | bash
npm create vagejs@latest my-site
cd my-site && npm install
claude

Prefer Homebrew? brew install --cask claude-code works too. See Anthropic's setup guide for Windows and other options.

How Claude Code finds the instructions

Claude Code loads CLAUDE.md from your project when a session starts. The starter's CLAUDE.md holds one line, @AGENTS.md, which pulls in the full vagejs instructions. Recent versions of Claude Code can also read AGENTS.md directly; Anthropic's docs recommend keeping the import anyway, because it covers older versions and never loads the file twice.

Want rules only for yourself, like your writing style? Put them in CLAUDE.local.md, which stays out of git. Anthropic's memory docs explain every location Claude Code reads.

Allow the safe commands

By default Claude Code asks before each command. Save yourself the clicks for commands that can't publish anything by adding them to .claude/settings.json:

{
  "permissions": {
    "allow": ["Bash(npm run check)", "Bash(npm run build)", "Bash(npx vage info --json)", "Bash(npx vage check --json)"]
  }
}

Leave vage deploy, deploy:github and indexnow off the list. vagejs asks before publishing anyway, and you want to be the one answering.

Prompts to try

Paste these into a Claude Code session:

Read AGENTS.md, then add an "About" page from the notes below. Ask me for anything the notes don't cover. Run npm run check until it reports 0 errors.

Run npx vage check --json and fix every error. For each warning you can't fix, tell me why.

Plan a blog post about our spring opening hours. Show me the title, description and outline first. Don't write the page until I approve.

Use plan mode (Shift+Tab) for bigger changes: Claude Code shows its plan before touching a file. More prompts are in the prompt library.

Good to know

• /memory shows which instruction files are loaded right now.
• If Claude Code seems to ignore a vagejs rule, ask it to re-read AGENTS.md. Instructions are guidance, not hard limits, which is why vage check exists.
• The site's own .github/workflows/ci.yml runs the same checks on every push, so mistakes get caught even when nobody runs them locally.