The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and synthesize information from untrusted external sources via web search tools.
Ingestion points: Uses searxng, duckduckgo, and perplexity MCP tools to fetch live web content (SKILL.md).
Boundary markers: The skill explicitly frames queries as falsifiable and defines acceptance criteria for sources to mitigate accidental obedience to embedded instructions.
Capability inventory: The skill includes instructions to generate "verification hooks" or shell commands to be executed in the local environment (SKILL.md).
Sanitization: Employs a technical rubric (references/source-quality.md) that prioritizes official documentation and maintainer-authored posts over untrusted sources like random gists or SEO blogs.
[COMMAND_EXECUTION]: The skill protocol instructs the agent to provide concrete commands or tests to verify claims locally. If these commands are derived from untrusted research data without sufficient validation, they could potentially include harmful operations.
[DYNAMIC_EXECUTION]: The "Implementation notes" and "Verification hooks" sections involve the generation of executable content (commands/tests) at runtime based on synthesized research data.

deep-research