The Conversational SDLC | Joseph Bironas | Medium

As we bring LLMs into the product development process it isn't just about the engineers doing things differently. All that context in a discovery session is critical for the LLM. How we write and think about features will need to evolve. This article highlights a first look at what that may look like.

The magic isn't the gates; it's the friction they re-introduce in just the right spots, and the context references they produce along the way.

• Less ambiguity up front. Writing the prompt contract feels bureaucratic until the first regen loop flies by unnoticed, and the model starts giving subtle clues that it no longer understands what needs to be done because I forgot to specify the prompt context correctly in the moment.
• **Interruption-aware iterations.**Cognitive-science studies and developer surveys agree that fragmentation disrupts flow more than interruptions themselves (blog.stackblitz.com, kjl.name). Capping loops keeps the brain in one piece.
• Reviewable reasoning. A reviewer who can read the model's chain-of-thought is far less likely to rubber-stamp copy-paste patterns.
• Post-mortem super-powers. Every production incident becomes replayable: prompts, responses, timestamps. Blameless and searchable.

The big thing I’m seeing repeatedly is including the prompt assets and other inputs into the repository right alongside the code itself. This is probably a very good thing in general.