The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/compress.py executes the claude command-line interface using subprocess.run(). This is used as a fallback for authentication when the ANTHROPIC_API_KEY environment variable is not set. The command execution uses a fixed argument list (['claude', '--print']) and passes file content through standard input (stdin), which minimizes risk of shell injection but still involves running a local executable.
[DATA_EXFILTRATION]: The core functionality of the skill involves reading local files and sending their entire content to the Anthropic API for compression. While this is the intended purpose, it constitutes data transmission to a third-party service. The skill attempts to mitigate risk by implementing a is_sensitive_path heuristic that blocks files with names or paths associated with secrets (e.g., .env, id_rsa, .aws/credentials).
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it interpolates raw, untrusted content from local files directly into the LLM prompt in scripts/compress.py.
Ingestion points: Reads files provided by the user via Path.read_text() in scripts/compress.py and scripts/detect.py.
Boundary markers: The prompt uses a simple header (TEXT:) but lacks robust XML-style delimiters or clear instructions for the model to treat the following content as data rather than instructions.
Capability inventory: The skill has the ability to write to the file system (Path.write_text()) and execute shell commands (subprocess.run()).
Sanitization: It performs a filename-based sensitive data check but does not sanitize or escape the content of the file before it is sent to the LLM.

caveman-compress