The Agent Skills Directory

[SAFE]: The skill performs localized text edits on markdown files within a specified workspace. It does not perform network operations, access sensitive system credentials, or request elevated privileges. Actual behavior aligns with its stated purpose of de-slotting and harmonizing paper voice.- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted markdown data from 'sections/*.md' and 'WRITER_SELFLOOP_TODO.md'. Evidence Chain: 1. Ingestion points: The agent reads markdown drafts and a TODO report from the local file system. 2. Boundary markers: The skill includes specific 'Role cards' like the 'Evidence Steward' to prevent content drift and maintain claims. 3. Capability inventory: The agent is authorized to update existing markdown files and create a status marker file. 4. Sanitization: The instructions provide explicit constraints to keep citation keys and factual meaning unchanged during the rewrite process.- [DYNAMIC_EXECUTION]: The included certificate script ('scripts/run.py') dynamically resolves the repository root to import project-specific helper functions ('tooling.common' and 'tooling.quality_checks'). This is a standard utility pattern for workspace-local tools and does not involve remote code execution or untrusted path interpolation.

style-harmonizer