The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill parses and executes instructions provided within the project's root CLAUDE.md file, particularly within a ## qa-full section. It is instructed to "honor whatever it states" regarding mandatory checks and command definitions.
Ingestion points: Root CLAUDE.md file, package.json, Makefile, and pyproject.toml (processed in the 'Project configuration' and 'Step 2' sections).
Boundary markers: Absent; the skill is directed to interpret plain prose instructions as configuration.
Capability inventory: Execution of arbitrary shell commands using allowed-tools such as Bash and Task.
Sanitization: No evidence of sanitization or validation of the strings retrieved from project files before they are executed as commands.
[COMMAND_EXECUTION]: The skill is designed to automatically identify and execute build and test commands from the target repository's environment.
Evidence: Under Step 2 (Tests & build), the skill is instructed to "Find the project's test/build commands from CLAUDE.md; if absent, infer from package.json scripts, Makefile, pyproject.toml, etc." and subsequently "Run the test suite and the build."
[DYNAMIC_EXECUTION]: The skill determines which tools and sub-skills to invoke (e.g., /defense, /db-optimize, /web-perf, /qa-only) based on the content of the branch diff and project metadata at runtime.
Evidence: The 'Related commands' and Phase-4 accounting sections describe a dynamic trigger matrix where various execution paths are activated based on file extensions and directory patterns found in the repository.

qa-full