The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to call the local claude CLI as a fallback mechanism for compression when an API key is not provided. This is performed using a fixed argument list and shell=False, which prevents shell command injection. The execution is limited to the claude --print command for processing file content.
[DATA_EXFILTRATION]: The skill transmits the content of user-specified files to the Anthropic API. To protect user privacy, it implements a comprehensive is_sensitive_path heuristic in scripts/compress.py that blocks files with sensitive names (e.g., .env, credentials, secrets) or those located in private directories (e.g., .ssh, .aws, .kube). Additionally, it enforces a 500KB file size limit to prevent excessive data transfer.
[PROMPT_INJECTION]: The skill processes untrusted natural language data from local files, creating an indirect prompt injection surface where embedded instructions in a file might influence the compression behavior.
Ingestion points: Local file content is read into memory via filepath.read_text() in scripts/compress.py before being passed to the LLM.
Boundary markers: The prompt template uses a 'TEXT:' label and 'STRICT RULES' to guide the LLM, though it lacks specialized delimiters to isolate untrusted content from the system instructions.
Capability inventory: The skill is capable of overwriting local files (after a backup is created) and executing the claude CLI.
Sanitization: The skill employs a robust denylist for sensitive file patterns to ensure private data is not processed by the LLM.

caveman-compress