The Agent Skills Directory

Instruction Override Patterns: The skill utilizes tags such as <system-reminder> and includes directives to 'ignore any existing tests' or 'ignore other tests entirely'. These patterns are designed to refocus the agent's behavior during a debugging session, but they resemble techniques used to override operational constraints.
Command and Dynamic Execution: The instructions provide specific code patterns for executing external shell scripts and starting background processes using Python's subprocess.run and Node.js's exec. While these are necessary for end-to-end testing, they involve the execution of local system commands based on the debugging context.
Indirect Prompt Injection Surface: The skill directs the agent to 'read logs' and 'analyze the output' of the application being tested. This creates a data ingestion point where content from the application's logs—which might include untrusted or external data—is brought into the agent's context, potentially influencing its subsequent actions.
Authentication Handling: The skill explicitly instructs the agent to ask the user for credentials if authentication is required. This is a secure alternative to hardcoding secrets, though it involves the agent handling sensitive information during the session.

creating-debug-tests-and-iterating