She stood at the door, laptop in hand, eyes a little nervous.
She had no engineering background, assigned to a live financial SaaS project. Not a tutorial, not a side project — a real system with real users. Approval workflows, budget management, annual forecasting. Every page backed by real business logic.
I was her mentor. I'd managed teams before, but those were experienced engineers who could run with a direction. This was different: taking someone with zero engineering background and guiding them to independent delivery at another company.
And I wasn't going to teach her traditional coding.
We met every two weeks. I wouldn't sit beside her writing code — instead, each session I'd review her progress, give direction, and teach thinking frameworks. The two weeks in between, she'd figure things out on her own, working with AI.
Teaching Thinking, Not Coding
Her first assignment wasn't writing code.
"Write down everything the company needs you to do. Every workflow, every process. Don't worry about format — just write it all down, as detailed as possible. Then use AI to organize it until it makes sense to you."
She looked at me like I'd asked her to write an essay instead of code. But she started writing.
This was my teaching strategy — Vibe Coding. No for-loops, no state management, no CSS. Think like a product manager: What problem does this system solve? How will users interact with it? Where does data come from, where does it go? Once you've thought it through, describe the requirements in plain language. AI turns it into code.
Code is AI's job. Understanding requirements, designing flows, judging logic — that's human work.
First review: 3 out of 5. Not high, but expected. Her PRD was logically clear, she asked the right questions. That matters a hundred times more than knowing React component patterns.
Five Whys
Second session. She brought a bug that had frustrated her for two weeks.
An approval list page should show pending items. It showed zero. Every time. She'd tried everything — telling AI "fix this list," "try a different approach," "reload the data." AI complied each time. Still zero.
The problem wasn't AI's code. It was the direction she gave AI. She was spinning at layer one.
I asked her one question: "Why does the page show zero?" Then kept going — empty array, API returned nothing, permission filter too strict, department mapping wrong, field names inconsistent from the original data import.
Five whys, layer by layer, until we hit bedrock. The root cause was buried in the moment the data was first imported.
Vibe Coding doesn't mean you don't need to understand the system. You can let AI write code, but AI won't ask "why" for you. Get the direction right first — then AI can help.
She developed a habit after that: before telling AI what to fix, write five "whys" first.
Draw the Target Before Shooting
Same bug. I did something that confused her — I wrote test cases before writing the fix.
"But the bug isn't fixed yet. What's the point?"
"Draw the target first, then shoot. Don't shoot the arrow and draw the target around wherever it lands."
In Vibe Coding terms: tell AI "after you're done, these scenarios need to pass," then let AI implement. If you don't define what "correct" means first, you can't judge whether AI's output is right.
She thought for a moment: "So tests are like acceptance criteria between me and AI."
Exactly.
The Blank Screen
One session, I was the one who broke things. Pushed a dashboard fix — entire page went white. Nothing rendered.
She watched. The mentor who'd just taught debugging had personally blown up the page.
I didn't panic, didn't make excuses. Opened the tools, spent forty minutes finding the cause, fixed it, added tests.
She told me later that session taught her more than all the previous ones combined. Not technique — attitude. Break something, fix it, add a safety net. No drama needed.
What you demonstrate matters more than what you teach.
"Am I Too Stupid?"
About three months in, she sent me a message.
"Why am I always behind schedule? Am I too stupid for this?"
I replied:
"Business logic is 10x harder than code. Even the mentor needs time to figure it out — it's not just you."
This wasn't comfort. It was fact.
What blocked her wasn't syntax — AI handled that. It was financial approval workflows, budget classification hierarchies, multi-layered permission logic. AI can write code for you, but AI won't understand the business for you.
The bottleneck in Vibe Coding is never the code. It's how deeply you understand the domain. The more precisely you describe, the higher quality AI produces. And precise description requires actually understanding the thing.
After that, her questions changed completely. No more "how do I use this component" — now it was "under what conditions does this approval get rejected?"
The Step Nobody Taught Her
During one code review, I noticed a small change.
She'd proactively added service-account-key.json to .gitignore.
Nobody told her to. She recognized on her own that a credential file shouldn't be committed. Even with Vibe Coding, you still need judgment about what should and shouldn't happen. AI won't proactively think about security for you.
Becoming the Company's First Engineer
End of the third month. She independently completed the annual forecast write-back feature.
Not a tutorial exercise. A real feature for real users, affecting the company's actual financial reports. She defined the requirements herself, designed the data structure, used AI to implement front and back end, handled edge cases.
From requirements to delivery, she did it herself. I only reviewed at the end.
Score: 4.5 out of 5.
| Session | Score | Key Milestone |
|---|---|---|
| 1st | 3.0/5 | PRD writing, workflow documentation, AI-organized requirements |
| 2nd | 3.5/5 | Five Whys learned — direction matters before AI can help |
| 3rd | 3.5/5 | Acceptance criteria concept, phased planning |
| 4th | 3.5/5 | Backend API integration (AI-assisted), blank screen lesson |
| 5th | 4.0/5 | Business logic breakthrough, security awareness |
| 6th | 4.5/5 | Independent: requirements → AI implementation → full delivery |
Look at this path. She didn't learn "how to write React." She learned how to turn a vague business requirement into a precise instruction that AI can execute. That's the core skill of Vibe Coding.
What the Mentor Learned
After the project, I wrote the entire experience into a document. Over a thousand lines.
While writing, I realized — many things only crystallized because I had to teach them. Five Whys was pure instinct before; I'd never articulated it as a method. Define acceptance criteria first? I didn't always do it myself. The blank screen? Without someone watching, I'd have patched it and moved on.
Ten principles, discovered in hindsight:
- Small steps — One clear goal per two-week cycle
- Embrace errors — Mistakes are the best material, including the mentor's
- Requirements first — Think clearly about what, then let AI do how
- Define acceptance criteria — Can't judge "done" without defining "correct"
- Five Whys — AI won't ask why for you; diagnosis is human work
- Specific feedback — Don't say "looks good." Say "why this design?"
- Psychological safety — Make it safe to fail and ask questions
- Normalize difficulty — Hard doesn't mean stupid; business logic is genuinely hard
- Track growth — Scores every session make invisible progress visible
- Write it down — Teaching crystallizes knowledge; writing crystallizes teaching
The Real Ending
A few months after the project ended, she sent me a message.
Not about a bug. Not about tech. She'd decided to enroll in a formal engineering training program. She wanted to learn programming for real, deeply.
I paused. The nervous person who'd stood at the door three months ago was now voluntarily walking toward this path. Not because the company required it — because she'd discovered something during those three months: turning a vague idea into something real that people actually use fascinated her.
Some people learn to code and think "this is a job." Others learn and think "this is what I want to do." The difference isn't talent. It's whether that first experience was real enough.
Her first experience wasn't Hello World — it was a live financial system. Her first debugging wasn't a textbook exercise — it was a production bug affecting real users. Those experiences gave her more than skills. They gave her certainty: "I can do this, and I want to keep doing it."
I replied: "Go for it. You already know the most important thing — think clearly before you act. Everything else is details."
The biggest achievement of mentoring isn't teaching a skill. It's igniting the motivation to keep going. Skills can be learned. Once motivation dies, nothing can be learned at all.
I'm Young, a freelance full-stack AI developer based in Taiwan. I've managed teams before, but this was different — using Vibe Coding at another company to take someone with no engineering background to independent delivery of financial features. Three months later, she chose to pursue engineering as a career. No syntax taught. Just thinking, and a spark.