I was analyzing my Claude Code logs and I realized something: I build by asking questions, not by issuing commands.
In hindsight, this makes sense for two reasons:
First, the answers let me verify that it has the same model that I do, meaning it primes the context with the right approach. The code that flows in the rest of that context window sits on the right foundation.
Second, it helps prevent "cognitive debt". I understand the theory behind the code, even if I don't look at every line. When I read it later, I'm already oriented.
Here are the questions from my logs that make the biggest difference:
Before anything starts: "What are the bounded contexts, and what is the concern?"
This one question has caught more architectural mistakes than any code review. The agent often wants to put new functionality wherever seems convenient. Asking about bounded contexts nudges it to think about where the code actually belongs. I've had it change the home of a feature many times just from this question.
Before planning: "What does DRY, SoC, and the bounded contexts suggest we do here? What are the tradeoffs?"
I name the specific principles I want applied and ask for tradeoff analysis, not just a recommendation. The agent is very good at reasoning through tradeoffs when you explicitly ask for them. It's not good at volunteering that reasoning on its own.
During review: "Fix any issues, rerun to verify, repeat until clean. If you need my input, check the knowledge base first, then give me three options and record my decision."
This is the one that saved me the most time. The agent auto-fixes trivial stuff, checks whether I've already made a relevant decision, and only escalates genuinely novel questions. Once I decide something, it never asks again.
When something breaks: "Do you already know the root cause, or do you need to investigate?"
I never say "fix it." Diagnosis and treatment are separate steps. The agent will happily apply a patch that hides the symptom if you let it.
These aren't magic prompts. They're the same questions a good tech lead asks a junior developer. The difference is I ask them hundreds of times a day across parallel agents, and the structure scales.