A GenAI skeptic describes a workflow called 'Interactive Rubber-Ducking' where an LLM asks one question at a time to help clarify and challenge design ideas rather than generating solutions. The technique uses a specific prompt in Claude Code with a capable model (Opus) to produce two artifacts: a Q&A log and a spec summary. A real example is shown: designing a second-level cache for the Pongo/Emmett OSS projects, covering 15 rounds of Q&A that cover cache interface design, cascading configuration, key strategy, optimistic concurrency, transaction behavior, and error handling. The resulting spec includes a pluggable PongoCacheProvider interface, MaybePromise return types, cascading config from client to per-operation level, and a detailed implementation plan.
Sort: