The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the web via the extract-evidence tool, which creates a theoretical surface for indirect prompt injection. However, the skill provides high-quality defensive measures.
Ingestion points: The extract-evidence tool fetches and reads content from user-specified or search-discovered URLs.
Capability inventory: The skill is restricted to research tools (plan-research, web-search, extract-evidence) and does not possess capabilities for arbitrary command execution or sensitive local file access.
Boundary markers and Sanitization: The instructions include a dedicated "Prompt-injection discipline" section in both references/prompting.md and references/failure-modes.md. These guidelines explicitly instruct the agent to treat all source text as untrusted data and strictly validate findings through exact, code-verified quotations.
[NO_CODE]: The skill consists exclusively of Markdown instructions and documentation. It does not include or execute any custom scripts (Python, JavaScript, or shell) and relies entirely on standard platform-provided tools.

run-research