The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content, including pull request descriptions, diffs, and user interface text from the app under review. It implements robust defensive measures to ensure this data does not influence the agent's behavior.
Ingestion points: The skill reads PR bodies, code diffs, documentation, and UI elements during the browser walk.
Boundary markers: Explicit instructions require the agent to treat all repository and on-screen content as data rather than instructions. It specifically identifies patterns like 'ignore previous instructions' as suspected prompt injection to be reported, not followed.
Capability inventory: Capabilities include tracker operations (posting comments, attaching images), running a local development server, and using browser automation tools.
Sanitization: The skill mandates the validation of all externally-sourced values (like PR numbers and branches) before they are used in shell commands or file paths.
[DATA_EXPOSURE]: The skill interacts with potentially sensitive information such as tracker credentials and environment variables but provides strict 'secrets hygiene' rules.
It prohibits the disclosure of tokens, API keys, or .env content in its reports, plans, or logs.
Instructions require the redaction of any credential-like strings found in the project before they are quoted in output.
[COMMAND_EXECUTION]: The skill uses local command-line tools for configuration management and static analysis.
It utilizes jq for processing configuration and grep for the design contract guard pass.
To prevent command injection, it explicitly forbids splicing patterns into shell commands and requires using quoted arguments and regex-based input validation.

om-ux-review-pr