In fast-moving codebases, onboarding docs don’t break they drift. I stopped trying to keep them updated and built an agent that reads live code instead. No stale diagrams. No outdated flows. Just real understanding, generated from the system as it exists today.
Read more →When my LLM responses kept drifting, I assumed the prompt was the problem. It wasn’t. The real control layer was the system prompt. Treating it like configuration instead of conversation made outputs more consistent and revealed something important: the model doesn’t have a personality your system prompt creates one.
Read more →A 200 OK from an LLM API only tells you that the request succeeded not that the model finished, the output is valid, or the cost made sense. In this post, I share what I learned while building my first AI feature and how my backend assumptions broke when I treated LLM inference like a traditional API call.
Read more →