skill-forge

Harden agent instructions against prompt injection by running them through an adversarial judge panel that tests for rationalization escapes and build-breaking shortcuts before deployment. This ensures cold-start agent safety and productivity in your specific repository.

Adversarial testing for AI agent instruction files is a quality assurance process that uses a judge panel to identify ambiguities, missing context, and prompt injection vulnerabilities in documentation like CLAUDE.md or SKILL.md before deployment.

Test CLAUDE.md files for agent productivity and build safety by subjecting them to a deterministic loop of observation, inspection, and amendment, followed by a 3-tier promotion gate to ensure strict fidelity and usability standards are met.

You can use automated quality gates to evaluate AGENTS.md documentation by applying a 3-tier promotion gate that enforces strict fidelity and usability standards, ensuring the instructions are safe and effective for cold-start agents.

AI agents fail following cold-start instructions due to ambiguities, build-breaking shortcuts, or missing context in instruction files. Adversarial testing identifies these issues before deployment by rigorously evaluating instruction fidelity against repository-specific requirements.

Automate quality assurance for agent skill definitions by implementing a deterministic multi-lens evaluation loop that inspects instructions for rationalization escapes and promotes them through strict quality gates, yielding high-fidelity documentation safe for deployment.