The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/compress.py invokes the claude CLI tool using subprocess.run to perform compression when the Anthropic SDK is not available. The command uses a fixed argument list and passes user content via standard input to prevent shell injection.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes the content of user-specified natural language files by interpolating them into a prompt for the LLM. This creates a surface where malicious instructions within a file could attempt to influence the LLM's output during the compression process.\n
Ingestion points: The skill reads files provided via the /caveman-compress command.\n
Boundary markers: The prompt in scripts/compress.py uses a TEXT: label but lacks high-entropy delimiters or explicit instructions to the LLM to ignore embedded commands.\n
Capability inventory: The skill has permissions to read and write files and execute the local claude CLI.\n
Sanitization: While code blocks are masked to prevent modification, the prose content is not sanitized against prompt injection patterns.\n- [DYNAMIC_EXECUTION]: The orchestrator script utilizes subprocess.run to execute a local binary (claude) determined at runtime using shutil.which to handle communication with the compression service.

caveman-compress