castor

CASTOR is an engineering gate skill for AI agents, for work that has to hold up in the real world: application code, firmware, HDL and hardware design. Instead of a longer checklist it is a small set of decision rules: what counts as a defect, what counts as proof, and which sign-off gates an agent may not skip. It was previously published as ironcode.

Its six laws are claims need evidence, analyze the cause before the fix, spec before style, trust nothing unverified, overruns are defects, and reversible by default. Risk tier sets the depth, from mechanical to behavioral to consequential work such as auth, money, migrations, bootloaders and power stages, and a pattern only becomes a finding when the agent can state trigger, mechanism and consequence.

The gate runs plan, build, gate and debug, and its verdicts are PASS, CHANGES NEEDED or INCOMPLETE, qualified by what was and was not examined. References load on demand for verification, correctness, security, resource safety, data access and ship readiness, plus domain guides for firmware, hardware and HDL. It installs as a Claude Code or Codex CLI skill, and the README states plainly that a PASS is a review discipline, not a certification.