Key Takeaways

  • Notion engineering manager Ryan Nystrom treats AI coding agents like Codeex as junior engineers: demanding explanations and forcing them to defend their work.
  • When faced with complex tasks, especially around CI, Nystrom starts with “I literally don't know what I'm doing here. You need to explain this to me like I'm a 5-year-old.”
  • He intentionally pushes back, telling the AI "you're wrong" to force it to cite specific, hard evidence, not just provide a solution.
  • This "diva" prompting approach is essential for ensuring correctness and building understanding in nuanced technical domains.
  • Fast CI is a silent partner, allowing quick validation of AI-generated code and iterative refinement of solutions.

The Method: How to Be a Diva to Your AI

Many founders treat AI as a magic box. You ask for code, it spits it out, and you hope it works. Ryan Nystrom, an engineering manager at Notion, takes a different approach. When he's tackling complex CI tasks or unfamiliar systems with an AI coding agent like Codeex, he doesn't just ask for answers; he demands understanding and pushes for proof. His method isn't about being polite; it's about being right.

Step 1: Admit Ignorance, Force Understanding

Nystrom's first move is counterintuitive: he explicitly admits his confusion. He feeds the AI lines like, "I literally don't know what I'm doing here. You need to explain this to me like I'm a 5-year-old." This isn't just about getting a simple explanation; it forces the AI to build up its reasoning from first principles, giving Nystrom a solid foundation to understand what's happening under the hood. For tricky configurations or obscure errors, this initial step ensures the AI doesn't skip critical context, which can be the difference between a quick fix and a cascade of new bugs. He sums it up: “I'm in over my head. Like you got to like explain it like I'm a 5-year-old.”

Step 2: Push Back, Demand Evidence

Once the AI offers a solution or explanation, Nystrom doesn't just accept it. He actively challenges it. “I will just like be like you're wrong,” he explains. This isn't a power trip; it's a strategic move to stress-test the AI's output. By telling it "you're wrong," he forces the agent to "defend your argument," demanding cited, hard evidence. In technical work, especially where correctness is paramount, guessing isn't an option. Nystrom wants the AI to back up its suggestions with robust reasoning, not just plausible code. As he puts it, "I just need to see the evidence that if I push counter to what it has done that it can like back up with like good pointed reasons." Claire Vo, his conversation partner, notes a similar approach: "Escape all caps. No. That's what I do. interrupt and no steer the conversation. Little arrow button has never gotten so much work."

The goal is not just to get any code, but to get correct code that Nystrom understands well enough to trust and integrate. This demanding interaction helps build a shared understanding and ensures the AI isn't just hallucinating a plausible answer.

Where This Breaks Down

This "diva" prompting method isn't a silver bullet. For simple, well-defined tasks where the stakes are low, the overhead of demanding kindergarten-level explanations and rigorous defense might slow you down more than help. If you're just asking for a basic Python script to parse a CSV, this approach is overkill.

It also assumes you, the human, have enough foundational knowledge or intuition to evaluate the AI's defense. If you truly know nothing about a topic, you might struggle to discern whether the AI's "evidence" is legitimate or just a convincing fabrication. The AI can still be confidently wrong, and without some human ability to spot flaws or ask follow-up questions, even the most aggressive prompting can fall short. This method works best when you know generally what you're doing, but the specifics are a lot more nuanced and you need to get this right.

What to Do With This

Next time you're stuck on a non-trivial technical problem—say, debugging a finicky CI pipeline or configuring a new serverless function—open your AI coding agent. Start by explicitly admitting your ignorance using Nystrom's exact line: "I literally don't know what I'm doing here. You need to explain this to me like I'm a 5-year-old." Then, when it offers a solution, don't just copy and paste. Challenge its claims. Tell it, "You're wrong. Defend your argument and show me the evidence." Push it until you feel confident in its reasoning, not just its output. This week, pick one specific technical roadblock and apply this "diva" method to build deeper understanding and higher correctness.