The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to automatically detect and execute arbitrary shell commands found in project configuration files such as CLAUDE.md, package.json, and Makefile to perform testing, linting, and building.
[EXTERNAL_DOWNLOADS]: The instructions direct the agent to automatically resolve missing dependencies by running package managers like npm, bun, or pip based on project requirements.
[PROMPT_INJECTION]: The skill explicitly mandates a policy of zero human interaction ("Never ask the user anything") and provides instructions to override the interactive logic of secondary skills, bypassing standard human-in-the-loop safety protocols.
[COMMAND_EXECUTION]: The agent systematically probes the local filesystem at ~/.claude/skills/ to locate, read, and load other skill definitions to extend its functionality.
[PROMPT_INJECTION]: The skill exhibits vulnerability to indirect prompt injection where malicious instructions embedded in a project's implementation plan or metadata could be executed autonomously.
Ingestion points: Project configuration files (CLAUDE.md, package.json, Makefile) and the implementation plan provided via files or conversation context.
Boundary markers: Absent; the agent is instructed to follow the detected commands and plan steps directly as implementation directives.
Capability inventory: Includes filesystem modification (branching, file writes), execution of arbitrary shell commands, and the use of network-enabled package installers.
Sanitization: Relies on an automated 'Roast' phase to calculate a safety score, but does not perform sanitization of input data or project metadata.

vs-autopilot