The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to interact with external web applications and process their content via the agent-browser tool. This creates a vulnerability surface for indirect prompt injection, where malicious instructions could be embedded in the web data (HTML, console messages, API responses) being analyzed. The instructions explicitly identify these sources as untrusted input and provide guidance to minimize obedience to embedded data.
Ingestion points: Web page content, console logs, and API responses processed in references/02-black-box-loop-user-flows.md.
Boundary markers: The skill includes explicit warnings in references/01-session-charter-setup.md to treat all external page content as untrusted input.
Capability inventory: The skill utilizes CLI commands for browser interaction, snapshotting, and data recording.
Sanitization: The workflow emphasizes black-box testing and warns against using implementation source code to justify behavior, which acts as a logical barrier between untrusted data and agent execution.
[COMMAND_EXECUTION]: The workflow relies on the execution of the agent-browser CLI tool and its subcommands. While these are presented as patterns for legitimate testing, they involve direct interaction with the host environment. The skill includes a Command Version Policy in SKILL.md requiring the agent to verify the tool's version and help surface before execution, which serves as a safety check against command syntax changes or unexpected tool behavior.

agent-browser-dogfood-workflow