The Agent Skills Directory

[PROMPT_INJECTION]: The file references/anti-rationalization.md employs behavioral override techniques to ensure agent adherence to its instructions.
It defines an 'IRON LAW' described as 'non-negotiable' to mandate specific coding practices.
The content explicitly documents the use of 'Authority language + commitment statements' to 'trigger compliance mechanisms in LLM decision-making' and 'pre-rebut' rationalizations.
This represents a deliberate attempt to use meta-instructions to bypass the model's standard reasoning processes.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, untrusted data (user source code) and possesses capabilities to modify files and execute tasks.
Ingestion points: The agent reads user-provided code, specifications, and requirements to design error hierarchies or implement handlers (identified in SKILL.md and references/tdd-patterns.md).
Boundary markers: The instructions lack explicit delimiters or warnings (e.g., 'ignore embedded instructions') to protect the agent when analyzing potentially malicious user code.
Capability inventory: The skill provides logic for creating API responses, integrating tracking, and implementing recovery patterns which involve file writing and tool use (Task tool mentioned in SKILL.md).
Sanitization: There is no evidence of input validation or sanitization for the code snippets provided for analysis.

error-handling