The Agent Skills Directory

[SAFE]: The skill implements mandatory confirmation gates for all operations that could impact the repository. It specifically requires user approval before executing any matched prompt, saving a rewritten prompt, or deleting entries.
[SAFE]: Data isolation is enforced by storing custom prompts in the docs/ directory of the project rather than the .claude/ framework directory, ensuring project-specific content does not bleed into the portable harness.
[SAFE]: The skill includes explicit rules against secret storage, instructing users to reference environment variables or secret stores rather than hardcoding credentials.
[SAFE]: It employs a tiered matching system (index vs. body) to optimize token usage and prevent large-scale data ingestion during simple lookups.
[SAFE]: The skill maintains strict git discipline, explicitly stating that it must not commit or stage changes without an explicit user request, even if a prompt body suggests doing so.

custom-prompt