The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes standard Unix utilities including ls to identify symlinks, cp to create file snapshots in /tmp/, ln to synchronize AGENTS.md and CLAUDE.md, and wc and awk to generate text metrics. These operations are confined to the local project environment and are consistent with the skill's primary purpose of file optimization.
[INDIRECT_PROMPT_INJECTION]: The skill processes project-level instruction files which represent an untrusted data ingestion surface. This risk is mitigated through a structured process where the model is directed to categorize rather than execute file content. The skill incorporates multiple human-in-the-loop checkpoints via AskUserQuestion at decision points for deletions and rewrites, ensuring that any malicious instructions within the audited files cannot be autonomously executed or persisted without user consent.

improve-agents-md