The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download and install the agent-browser tool, a core requirement for its browser-based research capabilities. It specifically suggests using npx skills add https://github.com/vercel-labs/agent-browser or npm install -g agent-browser. As the source is a well-known and trusted organization, this is documented as a standard operational dependency.
[COMMAND_EXECUTION]: The shared patterns in the skill instruct the agent to execute various local shell and batch scripts for environment setup and profile discovery, such as ./skills/paw-mkt-setup/assets/scripts/tool-discovery.sh and chrome-profiles.sh. It also provides commands for agent-browser operations and global Node.js package management.
[DATA_EXFILTRATION]: The skill describes a workflow for managing authenticated browser sessions, which involves discovering local Chrome profiles and saving authentication states to files (e.g., my-auth.json). While these instructions include security best practices like using .gitignore to prevent secret exposure, the programmatic handling of plaintext session tokens represents a potential data exposure surface.
[PROMPT_INJECTION]: The skill is subject to indirect prompt injection via its 'Live Website URL Audit' capability. This feature utilizes agent-browser or WebFetch to ingest content from external URLs provided by the user. If the audited site contains malicious instructions, the agent could potentially process them as part of its analysis, though the skill does not include any explicit instructions to bypass its own safety constraints.
Ingestion points: Content fetched from public URLs via agent-browser or WebFetch during site audits.
Boundary markers: Not present in the provided instructions; the skill advises the agent to resolve and apply context throughout the session.
Capability inventory: Includes Node.js package installation, local script execution, and browser automation.
Sanitization: No specific content sanitization or instruction-filtering measures are described for ingested web data.

paw-mkt-referral