The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted code and prose from user-provided files to perform automated cleanup and refactoring, creating a surface for embedded instructions to influence agent behavior.
Ingestion points: The skill ingests arbitrary source code and documentation files (READMEs, comments, PR descriptions) as defined in the target file list.
Boundary markers: While it utilizes a "scoped file list" as a boundary, it lacks specific instructions to ignore or neutralize malicious prompt instructions that could be hidden within the text or code comments of those files.
Capability inventory: The skill uses shell commands for linting and analysis, and it explicitly delegates modification tasks to external agents like Claude Code or Codex, which are granted file-write and execution permissions.
Sanitization: The skill relies on regression testing and verification passes to ensure behavior preservation, rather than sanitizing the content of the files it reads for potential injection patterns.
[COMMAND_EXECUTION]: The skill provides instructions for executing various command-line tools for linting, code analysis, and status recording.
Evidence: The skill references the execution of ruff check, vulture, knip, tsc --noEmit, and git log to identify cleanup candidates and verify results in references/cleanup-passes.md. It also records runtime status via the omh runtime record command in SKILL.md.

omh-ai-slop-cleaner