The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the claude CLI and jq via bash scripts (scripts/consult-claude-model.sh) to generate and verify advisory reports. The execution includes strict parameters to isolate the sub-process and restrict its behavior.
[EXTERNAL_DOWNLOADS]: Relies on the availability of the claude CLI and jq utility. The claude CLI is an official tool for interacting with well-known AI services from Anthropic.
[SAFE]: Implements a robust 'hardened runner' that uses explicit tool whitelisting (Read, Grep, Glob) and blacklisting (Bash, Edit, WebFetch, etc.) to ensure the advisor model remains read-only and cannot exfiltrate data or modify the environment.
[SAFE]: Mitigates potential indirect prompt injection from project files through a combination of system-level tool restrictions and explicit behavioral instructions for the advisor model. The security controls follow these criteria:
Ingestion points: Project files are read via the Read, Grep, and Glob tools in the consult-claude-model.sh script.
Boundary markers: The advisory prompt is separated from system instructions using a clear --- ADVISORY REQUEST --- delimiter.
Capability inventory: The advisor model is technically barred from file writes, command execution, and network access through the claude CLI's --disallowedTools and --strict-mcp-config flags.
Sanitization: The skill uses a comprehensive set of 'Hard constraints' within the advisor's system prompt to prevent it from revealing secrets or attempting to implement project changes.

claude-advisor