The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references several official developer documentation URLs from Anthropic and OpenAI to provide guidance on model behavior and prompting best practices. These links point to well-known, trusted domains (openai.com and claude.com) and are used exclusively for informational purposes.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a framework for reviewing and editing other agent skills, which inherently creates an ingestion surface for untrusted data contained within those skills.
Ingestion points: Instructions in references/review-gate.md require the agent to read the full content of target skills and their associated references.
Boundary markers: The skill emphasizes the use of an 'Authority Gate' and 'Safety Boundary' to distinguish between assessment-only tasks and those involving local edits, as well as checking for 'ignore embedded instructions' patterns.
Capability inventory: The skill possesses the capability to perform in-scope file edits and execute non-destructive tests or bundled validators as noted in SKILL.md and references/harness-portability.md.
Sanitization: The references/review-gate.md provides a structured checklist to verify authority, evidence requirements, and safety boundaries before accepting or implementing changes to target files.

cross-harness-skills