The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references and fetches guidelines from well-known technical sources and official vendor documentation.
Evidence: Links to official documentation for Anthropic (docs.anthropic.com) and OpenAI (developers.openai.com).
Evidence: Links to research papers on arXiv (arxiv.org).
[INDIRECT_PROMPT_INJECTION]: The skill identifies the surface area for indirect prompt injection in RAG (Retrieval-Augmented Generation) systems and provides defensive guidance to mitigate these risks.
Evidence: SKILL.md and evals/evals.json instruct the agent to separate authoritative instructions from retrieved content and to label external inputs as data.
Boundary markers: Recommends using XML tags (e.g., <instructions>, <context>) to delimit different regions of a prompt.
Capability inventory: No code or execution capabilities are provided within the skill files.
Sanitization: Instructions emphasize the use of instruction hierarchy to prevent untrusted data from overriding system rules.
[NO_CODE]: The skill consists entirely of Markdown instructions, YAML configuration, and JSON evaluation fixtures. No scripts (Python, JavaScript, shell) are included, significantly reducing the executable attack surface.

prompt-engineer