The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill dynamically loads persona instructions from project-specific files like CLAUDE.md or AGENTS.md. A malicious actor who can modify these files within a repository could inject adversarial instructions to override the agent's behavior.
[INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves a subagent (Comment Reaper) that processes code comments and suggests refactoring actions (identified as RESHAPE flags). If codebase comments contain adversarial instructions, they could influence the subagent's recommendations, which the main agent then uses to modify the codebase.
Ingestion points: Project configuration files (CLAUDE.md, AGENTS.md) and all source code files or diffs within the user-defined scope.
Boundary markers: The skill defines a specific audit step (Step 2) to check the subagent's report and reject application code changes, though it later applies refactoring logic based on the subagent's flags.
Capability inventory: The skill includes file modification capabilities to perform refactoring tasks such as renaming symbols, extracting functions, and adding runtime assertions.
Sanitization: The skill relies on instructional auditing logic where the lead agent reviews the subagent's output for correctness before proceeding.

banish