The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a structured surface for processing untrusted data fetched from the web (such as article excerpts and titles) and interpolating it into a final research report.
Ingestion points: Data enters the system via JSON evidence files processed by scripts/iterate.py and is rendered into markdown by scripts/compile.py.
Boundary markers: The skill includes comprehensive documentation for the host agent (references/methodology.md and references/sources.md) that defines a five-axis rubric (Authority, Recency, Evidence, Bias, Corroboration) for evaluating external sources, which serves as a mitigation against malicious content.
Capability inventory: The skill is limited to local file read/write operations for state management and report synthesis; it does not perform its own network requests or execute dynamic code from the processed data.
Sanitization: The skill utilizes Pydantic models for strict schema validation of all research data and performs basic string sanitization (stripping) on external excerpts.
[COMMAND_EXECUTION]: The skill uses local Python scripts (plan.py, iterate.py, compile.py) for workflow management. These scripts are invoked by the agent to transition between research stages and maintain the state file (plan.json). The execution is limited to standard file system operations and data processing within the skill's own directory.

deep-research