The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides documentation for implementing a hook-based system that injects data into the agent's prompt during runtime.
Ingestion points: The guide explains how to use user prompts, file paths (via glob patterns), and file contents (via regex patterns) as inputs for skill activation logic.
Boundary markers: The documentation describes a process where 'stdout becomes context for Claude', which involves injecting system-level instructions directly into the prompt stream without explicit guidance on using isolation delimiters or sanitization for the injected data.
Capability inventory: The hooks described in the guide have the capability to read local file system contents and execute local TypeScript scripts to determine what information to inject into the prompt.
Sanitization: There is no mention of sanitizing or escaping the content derived from external files or user prompts before it is interpolated into the agent's instructions.
[COMMAND_EXECUTION]: The troubleshooting and setup sections of the skill provide instructions for executing shell commands to manage the skill environment.
Evidence: Instructions include 'chmod +x .claude/hooks/*.sh' to grant execution permissions to local shell wrappers.
Evidence: The guide suggests running 'npm install' within the hooks directory to manage dependencies.
Evidence: The testing sections demonstrate using 'npx tsx' to execute local TypeScript files, passing JSON payloads via shell pipes for manual verification of hook logic.

skill-developer