The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input text (drafts) in several scripts. In scripts/declaude_review.py, segments of this text are sent to external AI APIs as part of a review prompt. This represents a potential surface for indirect prompt injection if the processed data contains instructions intended to influence the behavior of the reviewing model.
[COMMAND_EXECUTION]: scripts/declaude_diff.py uses the subprocess.run module to execute git show for version comparison. This is a standard project management function and is used within the context of the skill's diffing features.
[EXTERNAL_DOWNLOADS]: The scripts/declaude_rank.py script downloads a pre-trained model from Hugging Face's official repository for sentence embedding tasks. This is an expected and well-documented behavior for the ranking functionality.
[DATA_EXFILTRATION]: scripts/declaude_review.py transmits text segments (headers and opening/closing sentences) to external AI providers (Google or Anthropic) for analysis. This is a core component of the structural review workflow and is documented as the intended operation.

declauding