DEV Community

Cover image for I Let AI Write 100% of My Code for a Week — Here's What Actually Happened
Teguh Coding
Teguh Coding

Posted on

I Let AI Write 100% of My Code for a Week — Here's What Actually Happened

#programming #ai #productivity #beginners

Last week, I ran an experiment.

For 7 days straight, I didn't write a single line of code manually. Every function, every config file, every bug fix — I let AI handle it.

Not as a flex. As an honest test.

Here's what happened — the good, the bad, and the parts nobody talks about.

The Rules

Before we start, here's what I committed to:

  • ✅ AI writes ALL code (no manual typing)
  • ✅ I can describe what I want in plain English
  • ✅ I can review, accept, or reject AI output
  • ❌ No manually editing AI-generated code
  • ❌ No copy-pasting from Stack Overflow or docs

Tools I used: terminal-based AI agents, Cursor, and ChatGPT as backup.

Day 1-2: "This Is Amazing"

The first two days were incredible. I felt like a 10x developer.

What I shipped:

  • REST API with 6 endpoints (Express + MongoDB)
  • JWT authentication
  • Input validation with Zod
  • Docker setup with docker-compose

All in about 4 hours of "work" — which was mostly me describing features and reviewing code.

I literally made coffee while my API was being built. The future felt real.

Day 3-4: "Wait, Something's Off"

Then things got... interesting.

I started noticing patterns:

🔴 Problem 1: The Code Worked But I Didn't Understand It

The AI generated a middleware for rate limiting. It worked perfectly. But when I tried to explain it to a colleague, I couldn't. I had shipped code I didn't understand.

That's not engineering. That's gambling.

🔴 Problem 2: Subtle Bugs That Pass Tests

The AI wrote a function to calculate subscription billing. All tests passed. But it was rounding currency to 3 decimal places instead of 2. That's a bug that costs real money — and I almost missed it.

🔴 Problem 3: Generic Solutions for Specific Problems

I asked for a caching layer. The AI gave me a perfectly fine Redis implementation. But my app only has 50 users. I needed an in-memory cache with a 5-minute TTL. The AI over-engineered it because it didn't understand my scale.

Day 5-6: "I Need to Change My Approach"

By day 5, I realized the experiment wasn't failing — my approach was wrong.

I switched strategies:

✅ New Rule: AI Proposes, I Decide

Instead of blindly accepting code, I started treating AI like a junior developer:

  • "Write this function" → "Write this function and explain your approach"
  • "Fix this bug" → "What are three possible causes for this bug?"
  • "Add caching" → "Given 50 concurrent users, what's the simplest caching approach?"

The code quality jumped immediately.

✅ New Rule: Understand Before Merging

If I couldn't explain every line to a rubber duck, I asked the AI to simplify or explain. No more mystery code in my codebase.

✅ New Rule: Context Is King

The more context I gave the AI, the better the output. Instead of:

"Build a login page"
Enter fullscreen mode Exit fullscreen mode

I started saying:

"Build a login page for a B2B SaaS app. Users are enterprise 
customers who expect SSO. We use Next.js 15, Tailwind, and 
our auth provider is Auth0. Keep it minimal — no fancy 
animations. Mobile-responsive."
Enter fullscreen mode Exit fullscreen mode

Night and day difference.

Day 7: The Verdict

Here's my honest scorecard:

Metric Score
Speed ⭐⭐⭐⭐⭐ (3-5x faster)
Code Quality ⭐⭐⭐ (good, not great)
Learning ⭐⭐ (dangerous if passive)
Debugging ⭐⭐⭐⭐ (surprisingly good)
Architecture Decisions ⭐⭐ (needs human judgment)
Boilerplate/CRUD ⭐⭐⭐⭐⭐ (AI's superpower)
Complex Business Logic ⭐⭐ (risky without review)

What AI Is GREAT At

  1. Boilerplate — CRUD endpoints, config files, Docker setups. Let AI do this 100% of the time.
  2. Debugging — Paste an error, get a fix. This is where copy-paste AI actually shines.
  3. Learning new frameworks — "Show me how routing works in Hono" is faster than reading docs.
  4. Tests — AI writes solid unit tests, especially when given the implementation first.
  5. Refactoring — "Make this function more readable" works surprisingly well.

What AI Is BAD At

  1. Architecture — It doesn't know your team, your scale, or your constraints.
  2. Business logic — It doesn't understand that rounding to 3 decimals loses you money.
  3. Security — It generates "working" auth code that might have subtle vulnerabilities.
  4. Knowing when to say no — Ask AI to add a feature, and it will. Even if that feature is a terrible idea.
  5. Context across files — Unless you're using an agent that can read your codebase (not just a chat window).

The Framework I Use Now

After this experiment, here's my personal rule:

🟢 AI WRITES  → Boilerplate, tests, configs, CRUD, docs
🟡 AI ASSISTS → Debugging, refactoring, learning, prototyping  
🔴 I WRITE    → Architecture, business logic, security, data models
Enter fullscreen mode Exit fullscreen mode

AI is a power tool, not autopilot. You still need to know how to drive.

The Real Lesson

The developers who will thrive aren't the ones who write the most code OR the ones who blindly delegate to AI.

It's the ones who know when to write and when to delegate.

That judgment — that human layer — is what makes a senior developer senior. And no AI can replace it. Yet.

TL;DR

  • Days 1-2: AI is magic, I'm a 10x developer
  • Days 3-4: Wait, I don't understand my own code and there are subtle bugs
  • Days 5-6: Changed approach — AI proposes, I decide
  • Day 7: AI is a power tool, not autopilot
  • The rule: Green (AI writes), Yellow (AI assists), Red (I write)

Have you tried going full AI for a period? What was your experience? I'd love to hear if you hit the same walls I did.

Top comments (0)