One prompt to add AgentGuard.

Adding AgentGuard has always been lightweight. But there's one part no SDK can remove: understanding someone else's application.

In a real codebase, AI calls live everywhere. A model call inside a service. An agent loop in another module. A tool call buried behind a helper. A streaming endpoint built differently from everything else.

Installing the SDK isn't the hard part. The hard part is knowing where tracing and controls belong across the whole app.

Developers already have something that's very good at exactly that: their coding agent. So instead of another integration guide to read, we packaged our integration know-how into an AI coding skill for Claude Code, Cursor, Codex, Copilot and others.

From your application's folder:

Install the AgentGuard AI skill from github.com/ActaClad/agentguard-skills and use it to add AgentGuard tracing and guardrails to this application following best practices.

• Maps your app. Finds every LLM call, agent framework and tool call, and tells you up front if a provider isn't covered.
• Wires AgentGuard in. Installs the SDK and instruments model and tool calls, so every agent run becomes one clean trace.
• Keeps your keys out of the chat. Creates .env, asks you to fill it in, and waits.
• Proves it works. Runs your app once with a harmless test message, pulls the trace, checks it against our baseline, and fixes gaps until everything passes.
• Recommends guardrails without forcing them. They start off. It suggests the ones that fit your app; you decide what to enable.
• Audits existing setups. If AgentGuard is already there, it finds the gaps, upgrades an old SDK and leaves working code alone.

You review the diff. You stay in control.

We didn't make AgentGuard easier by cutting features. We put the integration knowledge where the work now happens: inside your coding agent.

This points to a broader change in developer documentation.

Docs are no longer just something developers read. They're becoming context that coding agents act on. The skill doesn't replace our documentation; it points the agent to the README of the exact SDK version it installed, and adds the judgement docs leave out: the right order, the common traps, and a clear definition of done.

The skill is open source under Apache 2.0, so you can read exactly what it tells your agent to do.