The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data by reading the content of other skills to perform its audit.
Ingestion points: Reads the directory or file path provided in $ARGUMENTS, subsequently reading SKILL.md and associated resource files (scripts, references, examples).
Boundary markers: The workflow does not explicitly instruct the agent to use delimiters or 'ignore' instructions when reading the content of the skill being audited.
Capability inventory: The skill operates with unrestricted tool access (default), including Bash and file system modification tools (Write, Edit), which are used to apply improvements.
Sanitization: No explicit sanitization or validation of the ingested skill content is performed before processing.
[DYNAMIC_CONTEXT_INJECTION]: The skill's documentation and reference files (references/frontmatter-options.md) promote the use of the !command syntax to inject real-time data into a skill's context.
Evidence: The skill recommends replacing tool calls with inline injections such as !git log and !git diff. It also mentions injecting environment variables.
Context: These recommendations are presented as optimization techniques for developer productivity. The SKILL.md itself does not contain any active, unescaped !command sequences that would execute upon its own invocation.
[SAFE]: The skill uses environment variables (e.g., ${CLAUDE_PLUGIN_ROOT}) to locate its own reference files, which is a standard and safe practice for skill organization. The overall design prioritizes context efficiency and follows established quality standards.

repair-skill