The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text files through an LLM to humanize them, creating a surface for indirect prompt injection.
Ingestion points: User-provided files read via scripts/cli.py and processed in scripts/humanize.py.
Boundary markers: The LLM prompt in scripts/humanize.py includes a 'STRICT RULES' section instructing the model to preserve code blocks, URLs, and technical terms exactly.
Capability inventory: The skill can read and overwrite local files, make network requests to the Anthropic API, and execute the claude CLI via subprocess.
Sanitization: The detect.py script checks for sensitive file paths and content (API keys, private keys) and refuses to process them in LLM mode.
[EXTERNAL_DOWNLOADS]: The skill optionally downloads machine learning models from HuggingFace for text detection and surprisal analysis.
Evidence: scripts/fetch_detectors.py and scripts/detector.py use huggingface_hub and transformers to download models like Oxidane/tmr-ai-text-detector and desklib/ai-text-detector-v1.01.
Context: These downloads target a well-known service (HuggingFace) and are opt-in features.
[COMMAND_EXECUTION]: The skill uses a subprocess to invoke the claude CLI as a fallback for LLM processing.
Evidence: scripts/humanize.py contains subprocess.run(["claude", "--print"], ...).
Context: The command uses a fixed argument list and passes user data safely via stdin, preventing shell injection.

unslop-file