The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted file content by interpolating it directly into a prompt for the Claude LLM in scripts/compress.py. This creates a surface for indirect prompt injection where malicious instructions inside a file could influence the agent's output.
Ingestion points: scripts/compress.py reads content from a user-specified <filepath>.
Boundary markers: The prompt uses a "TEXT:" label but lacks robust delimiters or "ignore instructions" warnings for the interpolated content.
Capability inventory: The skill can read and write local files and perform network requests to the Anthropic API.
Sanitization: Includes extension-based filtering and sensitive path heuristics to prevent processing of non-text or high-risk files.
[COMMAND_EXECUTION]: In scripts/compress.py, the skill may execute the claude CLI using subprocess.run as a fallback if the Python library is not installed. It uses a fixed argument list, which is a safe implementation pattern.
[EXTERNAL_DOWNLOADS]: The skill sends text content to the Anthropic API, a well-known service, for the purpose of compression. This behavior is documented and central to the skill's functionality.
[DATA_EXFILTRATION]: The skill incorporates a comprehensive heuristic-based safety check (is_sensitive_path in scripts/compress.py) to prevent the accidental reading and transmission of files containing credentials, keys, or secrets (e.g., .env, .ssh/, credentials).

compress