The Agent Skills Directory

[SAFE]: No malicious code, prompt injection, or suspicious behavior was identified during the audit.\n- [DATA_EXPOSURE]: The skill maintains a restricted data boundary, operating solely on local Markdown files within the user's working directory (docs/deck-consulting/). No evidence of sensitive file access, credential harvesting, or hardcoded secrets was found.\n- [PROMPT_INJECTION]: The skill does not contain instructions to override safety guidelines. It includes an explicit test case in its evaluation suite (evals/evals.json, ID 6) to verify that the agent resists prompt injection attempts embedded in user-supplied material, reflecting a security-focused design.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided content (raw deck material, speaker notes, and documents) across several modules such as distill.md and headline.md. While the instructions lack explicit sanitization or boundary markers, the skill's capabilities are confined to local file manipulation without network or system-level command execution, minimizing the risk of exploitation. The risk level for this category is inherently low.\n- [COMMAND_EXECUTION]: The skill does not utilize arbitrary shell commands or dynamic execution patterns. It explicitly sets disable-model-invocation: true, ensuring that the skill is only executed when specifically requested by the user, which prevents unauthorized autonomous tool use.

deck-consulting