The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection due to its ability to ingest untrusted data from live URLs via browser automation and user-supplied strategy files. It lacks robust instructions for data sanitization or the use of boundary markers when processing this external content.
Ingestion points: Browser tool outputs from live website audits and user-provided brand context files located in the .pawbytes/ directory.
Boundary markers: No explicit delimiters or "ignore embedded instructions" warnings are defined for the ingested data.
Capability inventory: Browser automation, shell tool usage, and file system write access within the project workspace.
Sanitization: No input validation or content filtering is specified for external data sources.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for the agent to install and configure external tools from a trusted source.
Evidence: In references/shared-patterns.md, the skill directs the agent to install the agent-browser tool from the official Vercel Labs GitHub repository (github.com/vercel-labs/agent-browser).
[COMMAND_EXECUTION]: The skill includes shell commands for the agent to run during environment setup and for performing research tasks.
Evidence: The references/shared-patterns.md file contains bash and batch commands for tool discovery (e.g., ./skills/paw-mkt-setup/assets/scripts/tool-discovery.sh), browser profile management, and tool installation.

paw-mkt-guerrilla