The Agent Skills Directory

Indirect Prompt Injection Surface: The skill is designed to capture user-provided feedback and corrections, then persist them into various documentation and instruction files. This creates a pathway where untrusted user input could potentially influence the agent's future instructions or behavior if malicious content is disguised as a legitimate correction.
Ingestion points: User feedback regarding implementation errors, preferred approaches, or rejected patterns (SKILL.md).
Boundary markers: While a structured format is provided for learning entries, there are no explicit technical delimiters or 'ignore' instructions defined to separate user-provided content from the agent's logic in the target files.
Capability inventory: The workflow involves reading and writing to sensitive repository locations, including .github/learnings/, instruction files, and design documents (SKILL.md).
Sanitization: The skill relies on 'generalization' and 'validation' steps to filter feedback, but does not specify automated sanitization or escaping mechanisms for the input being recorded.

feedback-learning