The Agent Skills Directory

Indirect Prompt Injection Surface: The skill scans user-provided source code, which constitutes processing of untrusted data. \n
Ingestion points: C# source code files are read by the agent or scanned via shell tools as described in the workflow (SKILL.md Step 3).\n
Boundary markers: The skill does not currently define explicit delimiters or "ignore instructions" to separate user code content from the agent's internal analysis logic.\n
Capability inventory: The skill utilizes file reading and grep-based shell commands to identify specific code patterns.\n
Sanitization: Source code content is processed directly for pattern matching without intermediate sanitization.\n
Context: The risk of instruction injection is mitigated by the skill's focus on technical pattern matching rather than open-ended interpretation of the code's prose or comments.
Shell Command Execution: The workflow provides pre-defined grep and wc command sequences to identify anti-patterns. These commands are constrained to text searching and counting within the target codebase, consistent with static analysis functionality.
Local Reference Management: The skill resolves paths for its optimization references (e.g., references/async-patterns.md) from its own installation directory rather than the user workspace. This is a secure practice that ensures the analysis logic is sourced correctly from the skill's bundled content.

analyzing-dotnet-performance