Rule-based verification checks an AI agent's output against objective, yes/no criteria — lint passes, automated tests pass, type checks pass, or any other domain-specific pass/fail rule. Because the check is binary and mechanical, it can be automated and run by the agent itself before a human ever looks at the output.
This is the first of two verification layers, paired with taste-based-agent-verification (the second, subjective layer), needed before trusting an agent with the autonomy described in Step 2 – AI Builds in Parallel (Orchestrator Model) and ai-supervised-autonomy-step of the Boris Cherny's 4-Step AI Maturity Framework: before letting an agent work unsupervised, you need a way for it to check its own work.
Apply: Build a Claude Skill that runs your domain's objective pass/fail rules (lint, tests, type checks, schema validation, etc.) against agent output, so the agent can self-correct before a human reviews it.