How to Code with AI Without Crying (Too Much)

Last week I showed how to start a project right—structure, prompts, docs, sanity. That kind of thing.

Then I published a few Reddit posts, and later stumbled upon a video from YCombinator called How to Get the Most Out of Vibe Coding. The advice? Surprisingly close to how I already work — except their version has better lighting and fewer existential monologues.

At the core of it, vibe coding (or whatever you want to call it — AI coding, Zen coding, etc.) is not about sprinting. It’s about leading. It’s about debugging calmly, planning like an adult, and talking to your AI like a confused but talented intern.

You’re not “hacking together a thing.” You’re the CEO of a very tiny startup. And your first hire is a senior AI dev who works 24/7 and never asks for lunch.

So this week, I just want to show how I work after the project is already started — when bugs creep in, or new features need to be shipped. The real-life workflow.

Here’s how I code with AI, in real life:

1. I keep one active ChatGPT “project” (or any other “AI” you’re using) that contains all major documents: PRD, tech notes, etc.

2. When something new pops up (a bug, a feature), I explain it in plain language. Like I’m talking to a team.

3. First, I ask the AI (inside Cursor) to mirror the problem back to me. “What did you understand?” This helps me catch misunderstandings before they write a single line of code.

4. If the AI’s summary is off, I refine it. If it’s good, I ask: “What questions do you have to better understand this?”

5. Then I request 2–3 possible solutions, but no implementation yet. Exploration only.

6. Once I pick a direction, then we move to implementation. Slowly, piece by piece.

7. After that: commit to GitHub, document the change, log it in a changelog file.

8. Yes, I ask it to help write documentation too — so I don’t forget what the hell we did two weeks later.

It’s not about dumping tasks on AI and praying. It’s about treating it like a high-powered junior — it needs leadership, not micromanagement. It’s on you to be the steady hand here.

And yes, I still refer back to the original product spec. It evolves. Things shift. But it’s always there.

In case you missed it:

• Here’s the free guide a lot of you downloaded already

• And here’s the full version for the price of coffee and croissant — I’m considering expanding it with more use cases next week

And I guess it’s fair to mention: this newsletter now has over 2,800 people reading it. Every week, we figure out how to build faster, smarter, and with fewer freak-outs. Glad you’re here. Let’s keep going.

Know someone wrestling with their AI side-project? Forward this and spare them tonight’s 3 a.m. meltdown. 300+ shares last week—thank you!

I’d love your help:
What’s the single most annoying step after you start an AI project?
– Prompt tuning
– Hunting bugs
– Keeping docs in sync
– Something else?

Hit reply with one short sentence. I’ll compile the top pain-points and ship fixes in next week’s newsletter.

See you next week.
—Miron