I really like the core of this idea (haven't tested the tool yet) - solve context-breaking not by better prompting but with a validation step that lives after the result but before human review. Solves the "CAPLOCK PLEA" approach completely, if not token-efficiently. But nowadays an ounce of determinism is worth its weight in plutonium.