Prompting AI to Rethink Workflow Yields Partial Improvements After Verification

The author experimented with prompting an AI to redesign a translation workflow from scratch based on goals rather than the existing architecture. The AI proposed a new approach, which was then verified in a git worktree against the current solution. Data showed the new approach did not outperform the original but offered partial value. The author used a sidechat to understand the proposal, then implemented the beneficial parts. Three lessons emerged: asking AI to think outside the framework can generate novel ideas; AI proposals must be verified with data, not accepted blindly; and sidechats are useful for clarifying non-critical questions without polluting the main context.