Most people default to using AI as an answer engine. You hand it a problem, and the model obligingly generates prose or code matching the implicit bias of your prompt. It feels fast, but it creates a dangerous illusion of velocity—baking unexamined assumptions into your architecture that only surface at runtime or code review.
This site almost started with a mistake born of that exact trap. My initial impulse was to build a full custom CMS wired with MCP integration and automated workflow harnesses—a textbook case of over-engineering before writing a single post. But instead of letting the model generate that massive spec, I turned the prompt around and asked it to grill my design choices.
I instructed the model to generate pros, cons, and probing questions for every proposed dependency. Under interrogation, the heavy CMS machinery collapsed. The automated workflow harnesses vanished, leaving behind a lean, static MDX pipeline that actually shipped.
That experience revealed the real power of AI in the development loop. The model doesn't bring original domain wisdom; it acts as an interrogative mirror. The human supplies domain truth; the model supplies friction. By forcing you to defend your decisions out loud, it exposes where your mental model is thin.
That experience changed how I approach writing as well as engineering. It led directly to building a custom writing skill—a digital ghostwriter whose job isn't to write prose for me, but to interview me relentlessly until my raw thoughts, vignettes, and core claims are clear enough to stand on their own.
Making this work in practice requires actively counteracting default LLM sycophancy. If you simply ask for feedback, the model will offer polite compliments. A proper interview prompt must explicitly mandate friction: forbid praise, force decision choices down every branch, and prohibit generating final artifacts until every dependency is grounded.
Whether building systems or writing essays, the pattern holds: AI is at its weakest when used to automate away the work of thinking, and at its strongest when used to force premature grounding. When you stop asking for answers and start demanding interrogation, you stop shipping plausible fluff—and start building architectures and prose that actually hold up.