The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection Surface. The skill processes user-specified files and interpolates their content directly into LLM prompts for compression. This creates a surface where malicious instructions within a file could attempt to override the agent's primary task.
Ingestion points: scripts/compress.py reads content from files provided as arguments to the /caveman-compress command.
Boundary markers: The ingested content is appended to the prompt template under a TEXT: header without additional sanitization or instruction-isolation delimiters.
Capability inventory: The skill has permissions to write files (overwriting the original) and execute shell commands (via the claude CLI).
Sanitization: No sanitization or structural validation is performed on the natural language content before it is sent to the LLM.
[COMMAND_EXECUTION]: Local CLI Invocation. In scripts/compress.py, the skill utilizes subprocess.run to call the claude CLI as a fallback when an API key is not present. This interaction uses fixed arguments (--print) and passes data via standard input, minimizing shell injection risks.
[SAFE]: Sensitive File Protection. The skill implements an explicit blocklist in scripts/compress.py (is_sensitive_path) that prevents it from reading or processing files typically containing secrets, such as .env, .ssh keys, or AWS credentials, before any external API calls are made.

caveman-compress