The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The skill uses the dynamic context syntax (!command) to execute a shell script at load time to generate a table of contents. This runs automatically when the skill is accessed.
[DATA_EXFILTRATION]: The skill reads the global Claude configuration file ~/.claude/settings.json, which often contains sensitive user data, credentials, or session information. While it uses this for configuration management, accessing such files is a data exposure risk.
[COMMAND_EXECUTION]: The skill executes complex shell commands and Node.js code using bun -e to perform file system operations, verify installation integrity, and modify settings. It also references an external runner upmd for executing markdown blocks.
[INDIRECT_PROMPT_INJECTION]: The skill parses YAML frontmatter from agents and skills located in ~/.claude/agents/. The extracted values are used to construct file paths for further read operations.
Ingestion points: Reads skills: preloads from agent files in ~/.claude/agents/ and skill files in ~/.claude/skills/.
Boundary markers: None.
Capability inventory: Uses Bash for command execution, bun for logic, and file system write access to ~/.claude/settings.json.
Sanitization: Basic trimming and quote removal are performed, but there is no validation for path traversal sequences like ../ when constructing paths from external file content.

setup