The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script (scripts/validate_research.py) to perform automated citation checking. This script verifies that file references and line numbers cited in research results actually exist within the local repository.
[INDIRECT_PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it interpolates user-provided data ({topic}, {item_name}, {item_description}) into the instructions for sub-agents. However, this risk is mitigated by the skill's design, which includes an 'evaluator' sub-agent specifically tasked with verifying if research conclusions are legitimately backed by source code evidence.
[SAFE]: The skill implements strong security boundaries by instructing the main agent not to read the raw content of research files, which prevents potential malicious instructions in the research data from affecting the main agent's behavior. Additionally, sub-agents are assigned appropriate permission levels, such as the evaluator agent being restricted to read-only access.

deep-research