The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents tools designed to ingest untrusted data from external URLs and local files (e.g., PDFs). It specifically addresses this risk by providing a 'Post-filtering strategy stack' to mitigate noise and potential injection from raw web results.
Ingestion points: crw search and crw scrape methods in SKILL.md.
Boundary markers: Mentions truncated: true flags for MCP calls to signal when content is clipped.
Capability inventory: Includes subprocess execution (Python), file writing to .crw/ or /tmp/, and network operations via CLI, MCP, and REST APIs.
Sanitization: Recommends a multi-layer filtering approach using rank-based triage, regex-based keyword density checks, and LLM-based snippet verification.
[COMMAND_EXECUTION]: Recommends using the crw CLI and piping outputs into Python subprocesses for filtering and data transformation. This is a core part of the recommended 'dynamic-search' pattern.
[DYNAMIC_EXECUTION]: Instructions encourage the generation and execution of 'one-shot' Python scripts and shell heredocs to process large search results outside of the primary agent context to save tokens and improve reliability.
[EXTERNAL_DOWNLOADS]: References several legitimate development dependencies, including the crw, anthropic, and firecrawl Python packages. It also describes setting up a local search backend using Docker.

crw-best-practices