The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources, creating a potential attack surface for indirect prompt injection.
Ingestion points: Research Mode uses agent-browser to fetch content from Product Hunt, competitor blogs, LinkedIn, X/Twitter, and Reddit (see references/launch-type-assessment.md and references/shared-patterns.md).
Boundary markers: The Reference Lookup Protocol in SKILL.md implements basic scoping by only reading matched files, but no explicit delimiters or instructions are provided to the LLM to ignore potentially malicious content embedded in the fetched web data.
Capability inventory: The skill has access to agent-browser for network interactions and interactive browser sessions, as well as file-write capabilities for saving deliverables to the vendor-specific ./.pawbytes/ directory.
Sanitization: There is no evidence of sanitization or filtering for content retrieved via agent-browser before it is processed by the agent.
[EXTERNAL_DOWNLOADS]: The skill requires the agent-browser tool, which is downloaded from a trusted organization's repository (vercel-labs/agent-browser).
[COMMAND_EXECUTION]: The skill provides instructions for executing shell commands via npx and agent-browser for research and setup tasks. These operations use well-known tools and follow standard integration patterns for research-oriented marketing agents.

paw-mkt-launch