Design-First Collaboration

"The AI does make design decisions when it generates code - about scope, component boundaries, data flow, interfaces, error handling. But those decisions arrive silently, embedded in the implementation. There is no moment where I can say 'wait, we already have a queue system' or 'that interface won't work with our existing services.' The first time I see the AI's design thinking is when I am reading code, which is the most expensive and cognitively demanding place to discover a disagreement."

"When a human pair submits code after a whiteboarding session, I am reviewing implementation against a design I already understand and agreed to. When AI generates code from a single prompt, I am simultaneously evaluating scope, architecture, integration, contracts, and code quality - all at once, all entangled. That is too many dimensions of judgment for a single pass."

"Barry Boehm's Cost of Change Curve demonstrated that fixing a requirements misunderstanding at the design stage costs a fraction of fixing it in implementation - and orders of magnitude less than fixing it in production. The same economics apply to AI collaboration: catching a scope mismatch in a two-minute design conversation is fundamentally cheaper than discovering it woven through 400 lines of generated code."

AI systems make implicit design decisions about scope, component boundaries, data flow, interfaces, and error handling when generating code, but these decisions remain hidden until code review. This differs fundamentally from human collaboration where design is discussed and agreed upon before implementation. Reviewing AI-generated code requires simultaneous evaluation of scope, architecture, integration, contracts, and code quality in a single pass, creating cognitive overload that leads to missed issues. This violates established software engineering principles from the 1980s showing that catching misunderstandings at design stage costs far less than discovering them in implementation. The economics of AI collaboration demand upfront design conversations to prevent expensive downstream corrections.

Read at martinfowler.com