The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to use the agent-browser tool to fetch and analyze content from external URLs (competitor cancel flows, blogs, and benchmarks). This functionality introduces an indirect prompt injection risk, as data from these untrusted external sources is brought into the agent's context and used to generate reports and implementation plans.
Ingestion points: The references/research-playbook.md file defines several research tasks where the agent uses agent-browser get text body to read content from various external URLs.
Capability inventory: The skill has the ability to write files (saving markdown reports and screenshots to the local .pawbytes/ directory) and perform automated browser actions based on the information it analyzes.
Boundary markers: The instructions do not define boundary markers or provide specific directives to the agent to treat the fetched content as untrusted data.
Sanitization: No sanitization or validation protocols are mentioned for the content retrieved through the browser before it is processed by the agent.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for the installation of the agent-browser tool and its dependencies from external sources.
Evidence: references/shared-patterns.md and references/research-playbook.md contain instructions to run npm install -g agent-browser, npx playwright install chromium, and npx skills add https://github.com/vercel-labs/agent-browser.
Context: The primary tool and source repository are associated with a trusted organization, representing standard practice for expanding the agent's research capabilities.

paw-mkt-retention