The Agent Skills Directory

[SAFE]: The skill follows a human-in-the-loop design where all proposed changes are presented as a checklist. It explicitly prohibits the automated modification of project files.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to perform file discovery using standard utilities like rg, fd, and find. The execution is restricted to identifying prose files and correctly excludes sensitive directories such as node_modules, .git, and environment configurations.
[PROMPT_INJECTION]: The skill processes untrusted project prose, creating a surface for indirect prompt injection. Ingestion points: project documentation files (.md, .txt, etc.). Boundary markers: absent from subagent instructions. Capability inventory: Read, Write, Bash, and Agent tools. Sanitization: mitigated by a multi-agent adversarial review phase that verifies findings via Grep and a mandatory human review of the final output checklist.

trim-language