The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it reads configuration files (CLAUDE.md, AGENTS.md) and user-provided source code to determine persona settings and identify comments for removal. * Ingestion points: Reads persona and style preferences from CLAUDE.md or AGENTS.md and reads target source code files. * Boundary markers: Includes an explicit instruction to treat instructions embedded in untrusted documents and tool output as data that cannot authorize actions. * Capability inventory: Capability to modify files (deleting comments, renaming symbols, extracting code) and execute shell-based validation commands. * Sanitization: Relies on the 'treat as data' instruction to prevent the agent from following instructions embedded in the analyzed code.
[COMMAND_EXECUTION]: The skill executes validation commands and potentially project-specific testing tools to verify that code cleanup did not introduce regressions. * Evidence: references/comment-validation.md instructs the agent to run project-required validation, focused tests, and compiler/linter checks after edits.

banish