Agentic AI: Solid Upfront Design Pays Off Later

Adding a new Startup Roast feature to the solo project DumbQuestion.ai became a real-world demonstration of how disciplined, boring engineering decisions in a first version pay off later. A system originally built to handle short questions had to process much longer, structurally different startup pitches, and it did so smoothly because prompt limits were parametric rather than hardcoded, web search was already an isolated reusable capability, the rendering layer was modular, prompt-injection handling was configurable, and the model/persona evaluation harness was built for extension from the start.

Across five concrete challenges — prompt sizing, adding market-context search, supporting markdown output, tuning prompt-injection defenses, and re-testing models and personas — the common thread is that early architectural investment turned what could have been rescue missions into small, predictable iterations. That directly reduces both developer time and the token cost of using coding agents.

The underlying lesson for engineers is that even a one-person side project benefits from lightweight SDLC discipline — not enterprise-style approval gates, but genuine extension points and separation of concerns. As teams lean more on agentic AI coding tools, clean architecture doesn't just make development faster; it makes those AI agents themselves more token-efficient and effective.