The Agent Skills Directory

[SAFE]: The skill content is purely educational and instructional, providing a diagnostic decision tree and quality metrics for managing the informational environment of an AI agent. It does not include executable code or scripts.
[EXTERNAL_DOWNLOADS]: The skill references documentation and research from trusted and well-known sources including Anthropic's official website, IBM's research blog, and ArXiv for grounding purposes. It also links to the author's official GitHub repository for the Skill Graph project.
[INDIRECT_PROMPT_INJECTION]: The skill addresses the inherent vulnerability of LLMs to untrusted data by advocating for isolation, provenance, and signal-to-noise optimization. 1. Ingestion points: System prompt, persistent memory, retrieved facts, tool outputs, and task metadata. 2. Boundary markers: Explicitly recommends isolation and source-aware selection to distinguish between instructions and data. 3. Capability inventory: Tools restricted to Read, Grep, Bash, and Edit. 4. Sanitization: Proposes Frequent Intentional Compaction (FIC) and immediate summarization of raw tool results to mitigate context pollution.

context-engineering