The Agent Skills Directory

[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it systematically navigates and interprets content from arbitrary web pages to find issues.
Ingestion points: Untrusted data enters the agent context through page snapshots and browser console output retrieved via the agent-browser tool (specifically in SKILL.md during the Orientation and Exploration phases).
Boundary markers: The instructions lack explicit delimiters or instructions to the model to ignore embedded commands within the target application's UI or content.
Capability inventory: The skill possesses capabilities to execute shell commands (mkdir, cp), perform file system writes (output directory management), and conduct extensive browser interactions (clicking, typing, and video recording).
Sanitization: No sanitization or filtering logic is applied to the web content before the agent processes it to decide on subsequent actions.
[COMMAND_EXECUTION]: The skill relies on shell commands for environment initialization and tool orchestration.
Execution method: Invokes Bash to create directory structures and copy report templates.
Context: Orchestrates the agent-browser CLI to automate browser sessions, which involves executing commands with parameters derived from the target application's state.

dogfood