Vibe Coding #2: AI Tool Confusion
How applying basic PM frameworks solved my vibe coding paralysis
This is part of my vibe coding learning journey. I'm an ex-Amazon PM learning to build with AI tools. My first project: a Luxembourg tax calculator. Each post breaks down one concept so you can apply it today, whether you're following the series or landing here for the first time.
The Tutorials
I got my first vibe coding idea. I had my first PRD and MVP written, but I couldn’t move forward.
I'd watched tutorials on Cursor, Claude, ChatGPT, Loveable, Bolt. Each seemed promising. From idea to ship in minutes.
Yet there I was, laptop open, multiple browser windows staring back at me. Each tutorial prompted another one. “If you haven’t watched my setup video, go here first.” An hour later, I’d watched far more tutorials and understood less than when I started.
Installing Claude Code threw npm permission errors. Terminal commands from tutorials didn't work. I spent hours troubleshooting the setup.
I needed to unblock myself. I returned to foundational guides instead of chasing the next YouTube recommendation. I searched back to:
Peter Yang ‘s 12 Rules to Vibe Code Without Frustration. It was easy for me to digest, probably because of the writing language. The corporate PM background helped.
Karen Spinner & Karo (Product with Attitude) ‘s The Ultimate Vibecoding Guide From Builders Who’ve Shipped.
Tool confusion isn’t a technical gap. It’s a clarity problem.
I fed my question back to Claude and Perplexity. “With this PRD, what is the basic tool set?” and “Why?”
I understood probably 50% of it. That was enough to move on.
Almost Gave Up
Halfway through trying to understand the setup, I had a realization: I could probably solve the whole tax deduction problem just by Googling. I have domain expertise in tax rules. What am I actually trying to do?
I had to reframe the project entirely.
It’s a coding learning project. The domain expertise is my advantage for validating what I build, not a replacement for learning how to build.
Domain expertise doesn’t eliminate the need to learn. It gives you an edge in knowing whether what you built actually works.
The Question That Matters
I fed my MVP and PRD to Claude. Claude asked me to choose between React and Tailwind or Replit. I didn’t understand what this choice was for.
So I reframed the question back to the AI. “To meet my MVP, what setup would you recommend? Explain why.”
AI is cheap this way. You can ask 100 dumb questions until you feel confident. You don’t need to schedule meetings. You don’t need to worry about wasting someone’s time. You just ask.
Think of AI as your employee, and you are the boss. Often, bosses don’t necessarily understand everything, but they need to make decisions based on information they receive. When you don’t have enough information to decide, ask your employee. Do the same with AI.
AI Marketing Traps
There are plenty of posts about how everyone can learn to code. They show projects shipped in 5 minutes.
All the “5 minutes,” “ultimate,” “the only” keywords are marketing. Do not get discouraged when you don’t hit the benchmark. The core idea is to appreciate that the tools give you the possibility to start.
Think about how PMs work. You write PRDs. You sit in meetings convincing stakeholders. You secure resources. You ask developers to carve out time to build your ideas. That entire process exists to get permission to start.
AI tools shift you from “I need to convince XX, YY people first” to “I can start today.”
Three Amazon Frameworks That Solved This
After the installation chaos settled, I noticed what I’d actually been trying to do. I was applying frameworks I already knew from work. These aren't Amazon-specific. If you've ever organized a team project, delegated tasks, or decided whether to sleep on a decision, you've used versions of these too.
Framework 1: Create Swim Lanes (RACI for AI Tools)
When we mapped complex processes at work, we’d draw RACIs. Responsible. Accountable. Consulted. Informed. Who owns what. Who’s responsible for which part. Clear boundaries. No overlap.
I needed to apply the same thinking to AI tools.
Perplexity does the research. Claude Code does the building. Claude Chat handles strategy and helps my understanding.
One does facts. One does files. One does thinking. No overlap means no decision paralysis.
Framework 2: Recognize Two-Way Door Decisions
Some decisions are hard to reverse. Those need deep analysis. Some decisions are easy to undo. Those need speed, not perfection.
Most AI tool choices are two-way doors.
When Cursor prompted me to upgrade, I switched to Claude Code.
This is why I didn’t need perfect understanding before starting. The door swings both ways.
Framework 3: Decide Based On Calculated Risk
I understood 50% of what Claude recommended and started anyway.
React, Node, Vercel - these were one-off setups. I didn't need to deeply understand them. I just needed step-by-step instructions to follow. My goal wasn't to become an expert in dev environments. It was to get past setup so I could learn by building.
Setting up a dev environment? Follow instructions and get it done.
These Three Frameworks Complete Each Other
Swim lanes tell you which tool to use. Two-way doors tell you it’s okay to try without perfect clarity. Calculated risk tells you when your understanding is enough to move on.
I had a decision framework. Does this task need facts verified? Start with Perplexity. Does it need files created? Use Claude Code. Do I need to understand why? Ask Claude Chat.
Nothing Starts From Scratch
When I began learning to vibe code, I thought I’d be starting from zero. I catch myself thinking this every time I try something new. And each time, I realize: nothing actually starts from scratch. Skills transfer.
If you’re a PM trying to start building, you already know how to create swim lanes. You already know how to identify two-way door decisions. You already know how to calculate risk with incomplete information.
You just haven’t applied those skills to AI tools yet. That’s the gap. Not technical knowledge.
Key Takeaways
Most AI tool choices are two-way doors. Choosing React vs Replit, Cursor vs Claude Code, one tutorial approach vs another? These are reversible. You don't need deep analysis. You need to start and adjust.
Ask AI “Why?” like you’d ask an employee. When AI presents choices you don’t understand, reframe the question back: “What would you recommend and why?” You don’t need to understand everything before deciding. You need enough context to move forward.
Swim lanes for AI tools eliminate decision paralysis. Assign each tool one clear job: Perplexity for facts, Claude Code for files, Claude Chat for strategy. No overlap means you know which tool to open.
Ignore the “5 minutes” marketing trap. AI tools let you start today without convincing stakeholders, securing resources, or waiting for developer availability. The real value is the permission, not speed.
Use PM skills you already have. RACIs, two-way doors, and calculated risk aren’t new skills to learn. They’re existing skills to apply in a new context.
Partial understanding is enough to start. Learning by doing works, but tutorials help you understand what, why, and how to start. Don't watch tutorials to optimize from 20 minutes to 5 minutes. Watch to understand the minimum viable setup, then stop and build.
Voilà, thanks for being here,
Ting
Related posts:
It’s great to see your approach to vibe coding. Questioning AI's output is super useful. I do this a lot to learn about its reasoning behind technical choices. It also helps me flag when its approach doesn’t make sense.