The Agent Skills Directory

[SAFE]: The skill package is composed entirely of Markdown documentation, templates, and research references. It does not include any executable scripts, binaries, or platform-specific configuration files that could pose a direct security risk.
[SAFE]: All external URLs and academic references point to well-known technology vendors and research organizations, such as OpenAI, Anthropic, Sakana AI, and METR. These resources are documented neutrally for educational context and provenance.
[NO_CODE]: There are no software dependencies or code execution patterns present in the skill. The content serves as a conceptual framework for designing robust agent prompts.
[INDIRECT_PROMPT_INJECTION]: The skill describes a methodology for agents that ingest and process external data, which defines a surface for indirect prompt injection.
Ingestion points: Processes external research data, web search results, and artifacts passed between delegated subagents.
Boundary markers: The documentation explicitly advocates for the use of pseudo-formal task specifications and strict success predicates to prevent agents from being steered by untrusted content.
Capability inventory: The skill discusses architectures involving tool execution, filesystem access, and parallel agent orchestration.
Sanitization: The methodology emphasizes defensive measures such as JSON schema validation, adversarial auditing by fresh-context agents, and domain-specific failure-mode checklists.

long-horizon-prompting