The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements advanced defenses against indirect prompt injection (CWE-1427) by requiring the agent to treat all data from target repositories as untrusted and ignore any directives found within it. Explicit instructions are provided to delimit ingested content and maintain isolation between targets to prevent cross-contamination.
[DATA_EXFILTRATION]: Includes mandatory secrets-scanning of staged changes before any commit occurs, utilizing tools like gitleaks or detect-secrets to prevent data leakage. It explicitly forbids the use of WebFetch or WebSearch to bypass authenticated audit trails.
[COMMAND_EXECUTION]: Uses authorized CLI tools (gh and git) with human-in-the-loop validation. A mandatory Pilot stage ensures that any fanned-out change is first proven on a single target and reviewed line-by-line.
[CREDENTIALS_UNSAFE]: Strictly follows the principle of least privilege by reusing existing authentication sessions (gh) and prohibiting the injection of new secrets, tokens, or writing credentials to disk.
[SAFE]: The overall architecture incorporates high-friction safety gates, including deterministic branch naming and detailed reconciliation reports (selected = applied + skipped + failed), ensuring all bulk operations are traceable, reviewable, and reversible.

force-multiplier