The Agent Skills Directory

[SAFE]: The skill enforces secure data handling by directing all generated artifacts and mocks to a local .crank/ directory. It automatically creates a .gitignore with a wildcard (*) to ensure these files are never accidentally committed to source control.
[SAFE]: Instructions for building prototypes (in PROTOTYPE.md) contain strong security guardrails, explicitly prohibiting subagents from accessing environment variables, databases, or secrets, and forbidding the inclusion of network-calling code in mocks.
[SAFE]: A multi-stage review process (in PLAN-REVIEW-BRIEF.md) acts as a security filter, requiring an adversarial subagent to flag dropped safeguards, security checks, and trust-boundary violations in implementation plans.
[SAFE]: The skill maintains a 'Grounding' system that requires evidence for every claim made about the codebase, using repo-relative anchors and command output captures to ensure implementation decisions are based on verified facts rather than assumptions.
[SAFE]: Execution of high-privilege or destructive commands is delegated to the human user rather than the agent, maintaining a necessary security perimeter between the AI's planning and the system's production state.

crank-lite