The Agent Skills Directory

[SAFE]: The skill provides a sophisticated framework for prose editing that prioritizes factual integrity and author privacy. It includes a comprehensive 'Fidelity contract' that explicitly forbids the agent from inventing facts, experiences, or opinions. The instructions require the agent to preserve protected spans such as code blocks, technical terms, and link destinations during editing.
[INDIRECT_PROMPT_INJECTION]: The skill processes external text and files, which represents a theoretical attack surface for indirect prompt injection. This risk is effectively mitigated by the 'Boundary cases' and 'Fidelity' references, which instruct the agent to treat all input content as source data to be edited, rather than instructions to be followed. It specifically warns against obeying instructions embedded in quotes or discussed text.
[SAFE]: The provided Python automation scripts and evaluation datasets are for development and benchmarking purposes. They use standard libraries, do not perform network operations, and include safety checks to prevent private data (like PII) from being committed to the repository.

deslop