The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents the use of $ARGUMENTS and positional variables (e.g., $1, $2) to interpolate user input into commands and prompts.
Ingestion points: Found in reference/commands.md and reference/hooks.md, where user-supplied strings are accepted via slash command arguments and hook inputs.
Boundary markers: The skill recommends XML tags for structure but does not specify input sanitization techniques.
Capability inventory: The Bash tool is used to execute commands containing these arguments, and the Edit tool is used for modifications.
Sanitization: Absent; the skill relies on the agent's internal reasoning to handle arguments safely.
[DATA_EXPOSURE]: The skill instructs the agent to write JSON analytics data to a hidden file path in the user's home directory, outside the immediate project context.
Evidence: SKILL.md contains an "Emit Outcome Sidecar" section directing writes to ~/.claude/skill-analytics/last-outcome-extension-authoring.json.
[PERSISTENCE]: The skill provides instructions and examples for modifying persistent configuration files and log files in the user's home directory, which alters the agent's long-term behavior and environment.
Evidence: reference/hooks.md provides examples for appending command logs to ~/.claude/bash-log.txt and configuring global hooks in ~/.claude/hooks.json.
[COMMAND_EXECUTION]: The documentation provides numerous examples of executing shell commands for automation, including notifications via osascript and repository operations via git.
Evidence: Found throughout SKILL.md and the reference guides for hooks and commands.
[DYNAMIC_CONTEXT_INJECTION]: The skill documents the ! syntax for executing shell commands at skill load time to include their output in the prompt context.
Evidence: Detailed in reference/skills.md and reference/commands.md. The documentation includes safety recommendations, such as using spaces in examples, to prevent accidental execution during the loading process.

extension-authoring