The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its processing of untrusted external data.
Ingestion points: Public review text and Google Business Profile data fetched via the local/search and maps/search tools (referenced in SKILL.md).
Boundary markers: None explicitly defined in the instructions to separate external review text from the agent's instructions.
Capability inventory: The skill can read local files (.agents/product-marketing.md) and perform network operations via the UnifAPI MCP server.
Sanitization: No specific sanitization or filtering of the review text is mentioned before the agent analyzes it for 'neighborhood language' or 'velocity'.
[DATA_EXFILTRATION]: The skill accesses local files that may contain marketing configurations or product strategies.
Evidence: Instructions in SKILL.md (Workflow Step 1) direct the agent to read .agents/product-marketing.md and .claude/product-marketing.md if they exist. While these are likely intended for configuration, accessing hidden dot-files is a pattern associated with data collection.
[EXTERNAL_DOWNLOADS]: The skill provides installation instructions that involve executing code from an external repository.
Evidence: The README.md file suggests running npx skills add unifapi-agent/agents. This command downloads and executes code. However, the package belongs to the skill's official author ('unifapi-agent'), which aligns with standard vendor distribution practices.

agent-reputation-benchmark