The Agent Skills Directory

[SAFE]: The skill instructions focus on high-fidelity prose editing. It explicitly forbids the agent from inventing facts, dates, or citations, and provides a 'fidelity contract' (references/fidelity.md) to ensure semantic relationships and numbers are preserved.
[SAFE]: For file-editing operations, the skill mandates 'byte-for-byte' preservation of non-prose regions, including frontmatter (YAML/TOML/JSON), fenced code blocks, and Markdown link targets. This prevents the AI from accidentally corrupting configuration or injecting malicious links during text editing.
[SAFE]: The inclusion of an evaluation suite (automation/eval/) demonstrates a structured approach to testing. The Python scripts provided (check_cases.py and make_blind.py) are benign utilities for validating test data and do not perform network operations or access sensitive system files.
[SAFE]: The skill's 'voice calibration' feature includes explicit privacy boundaries, instructing the agent to use only user-supplied samples and forbidding the persistence of author profiles or searching the filesystem for additional data.
[SAFE]: Installation instructions utilize the 'npx' package runner to fetch the skill from the author's own repository. This is a standard and expected mechanism for skill deployment on the platform.

deslop