The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes dynamic context injection (!pwd, !find) and restricted shell commands (git status, ls, wc) to map the project structure and analyze file sizes. These operations are limited to discovery and do not execute untrusted external code or perform dangerous system modifications.
[PROMPT_INJECTION]: Although the skill is designed to process files containing natural language instructions (like CLAUDE.md), it mitigates potential indirect prompt injection by requiring explicit user approval through AskUserQuestion before any content is deleted, moved, or consolidated. This human-in-the-loop requirement prevents automated exploitation by malicious content residing in the files being audited.

meta-context-diet