The Agent Skills Directory

[Command Execution]: The skill provides instructions on using shell scripts for plugin hooks and Model Context Protocol (MCP) servers.
[Evidence]: Examples in advanced-plugin.md and standard-plugin.md use ${CLAUDE_PLUGIN_ROOT} to execute local scripts and binaries. This is a standard pattern for plugin functionality and promotes portability.
[Indirect Prompt Injection]: The skill outlines a pattern for using 'prompt' type hooks to analyze untrusted data like code or shell commands.
[Ingestion points]: External data enters the context via Write|Edit or Bash event matchers in hooks/hooks.json.
[Boundary markers]: The examples do not demonstrate the use of delimiters or 'ignore embedded instructions' warnings.
[Capability inventory]: The documented plugin architecture supports shell script execution (bash), file modifications, and network operations via MCP tools.
[Sanitization]: While some examples show basic data sanitization (e.g., sanitize(data) in standard-plugin.md), the core prompt-based evaluation depends on the model's instruction-following capabilities.
[External Services]: The documentation describes integration with various external platforms and monitoring tools.
[Evidence]: Examples include references to GitHub, Slack, Datadog, and Kubernetes. These integrations are part of the intended advanced functionality for enterprise-grade plugins.

Plugin Structure