The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the Bash tool to interact with its core logic, specifically by calling python3 "${CLAUDE_PLUGIN_ROOT}/core/nutmeg.py". This environment variable refers to the local installation path of the toolset.
[DYNAMIC_EXECUTION]: The agent is instructed to create analysis scripts in Python, R, or SQL and execute them via the nutmeg run command. While the skill mandates a 'gate' mechanism where the user must approve the code before execution, this pattern involves the runtime generation and execution of arbitrary code.
[DATA_EXFILTRATION]: The analysis execution command (nutmeg run) supports a --sends argument (e.g., --sends api.example.com:player_id,columns) designed to transmit data to external hosts. This facilitates the export of potentially sensitive research data to third-party services.
[EXTERNAL_DOWNLOADS]: The skill documentation explicitly mentions ingesting data from 'a scrape, a download or an unknown source,' which involves fetching external content into the local environment for processing.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data (events, provider facts, etc.) to generate reports and claims. There is an inherent risk of indirect prompt injection if the ingested data contains malicious instructions designed to manipulate the agent's analysis or output, although the structured 'claim ledger' and human review steps serve as significant mitigations.

nutmeg-research