Use vagejs with your AI coding agent
Your assistant knows the rules.
Every new vagejs site comes with written instructions for AI coding assistants, a command that describes the whole tool as data, and checks that tell the assistant exactly what to fix. Your assistant spends its time on your content instead of guessing how the framework works.
What your assistant gets
AGENTS.mdin the project folder: the commands, recipes for common jobs (add a page, rename one, add a blog post) and the rules, such as "never invent dates or authors". Most coding agents read this file on their own.CLAUDE.md, which contains a single line,@AGENTS.md, so Claude Code loads the same instructions.npx vage info --json: every command, option and setting, with JSON Schemas for the config and frontmatter. Assistants are told to use it instead of guessing.--jsonoutput frombuild,check,auditand the deploy commands, where every problem names the file to fix.- Publishing that waits for you. Deploys, pushes and IndexNow submissions show their plan and stop for a yes, so an assistant can't publish your site on its own.
Which file each tool reads
| Tool | Reads on its own | Setup needed |
|---|---|---|
| Claude Code | CLAUDE.md, and AGENTS.md in recent versions |
None: the starter's CLAUDE.md imports AGENTS.md |
| OpenAI Codex | AGENTS.md |
None |
| Cursor | AGENTS.md and .cursor/rules/*.mdc |
None |
| GitHub Copilot | AGENTS.md and .github/copilot-instructions.md |
None in VS Code; the cloud agent needs a setup workflow |
| Gemini CLI | GEMINI.md |
One setting to add AGENTS.md |
| Other agents | Varies | Devin Desktop, Cline, Aider, Zed and more |
The loop every assistant follows
npx vage info --json # learn the commands and settings
# edit pages in src/pages/
npx vage check --json # fix every error it reports
npx vage build --json # build the finished siteThe instructions tell the assistant to run check after every change and to finish only when it reports zero errors. Warnings it can't fix get reported to you, not hidden.
Rules worth keeping
- Let it run the safe commands freely.
npm run check,npm run buildandnpx vage infoonly read your project or writedist/. Each guide shows how to allow them without a prompt every time. - Keep publishing manual. Don't put
vage deploy,deploy:githuborindexnowon an auto-approve list, and don't use "approve everything" modes on a project you care about. - Facts come from you. The instructions tell assistants to ask you for titles, dates, authors and claims instead of inventing them. If yours makes something up, point it back to
AGENTS.md. - Keep secrets out of chat. API keys belong in environment variables or
vage connect, never in a prompt or a page.
Ready for something to paste? The prompt library has prompts for everyday jobs that work with any of these tools.