Your experience with Hypershell points to an interesting possibility: LLMs as DSL translators rather than replacements. This could actually democratize DSLs by lowering the learning curve while preserving their domain-specific benefits. The real opportunity might be DSLs optimized for both human semantics and machine translation.