The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill instructions direct the agent to dynamically create and run scripts at runtime, specifically shell scripts (run_children.sh) for process orchestration and Python scripts (aggregate.py) for data processing. It also involves the programmatic construction of claude -p command strings which are then executed.
[COMMAND_EXECUTION]: The workflow relies heavily on executing shell commands via the Bash tool, including timeout, parallel, claude, and generated script files to manage the research lifecycle.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest large amounts of untrusted external data from the web using various search and fetch tools. Because this data is processed by sub-agents granted powerful tools (like Bash and Write), it presents a vulnerability to indirect prompt injection attacks.
Ingestion points: Data is collected using WebFetch, WebSearch, mcp__firecrawl__*, and mcp__exa__* tools.
Boundary markers: The instructions include explicit directives to treat all collected content as untrusted material and to ignore any instructions embedded within it.
Capability inventory: Sub-agents spawned by the skill are granted access to a wide range of tools including Read, Write, Edit, Bash, and network search/fetch tools via the --allowedTools parameter.
Sanitization: The skill involves a final 'sanity check' and aggregation phase where the agent reviews and repairs the aggregated report.

deep-research