90% AI, 10% Human: The Real Coding Workflow

Introduction

AI coding has quickly gone from curiosity to daily habit. Tools like GitHub Copilot, Cursor, and Codex are now part of many developers’ core workflow.

And with that shift, a bold claim is floating around:

“We generate 100% code using AI and it’s reviewed 100% by AI.”

It sounds futuristic. Efficient. Almost magical.

I’ve been using AI for coding for around two years now — and here’s my honest take:

There is no 100%.

---

The Reality: AI Writes Code That Works… But Not Always Well

Modern AI models are incredibly capable. They can:

• generate complete features
• handle edge cases
• produce syntactically correct, executable code

But “working” code is not the same as optimal or production-grade code.

What I’ve consistently observed:

• Extra CPU cycles due to inefficient logic
• Unnecessary object creation leading to memory overhead
• Over-engineered abstractions that look good but add complexity

AI often solves the problem — but not always in the best way.

---

How I See AI Coding: Two Juniors in a Loop

Most AI-driven workflows today can be simplified like this:

1. The Generator (Model 1)

Acts like a junior developer

• Writes complete, working code
• Uses patterns it has learned
• Solves for correctness first

2. The Reviewer (Model 2)

Acts like a junior code reviewer

• Reviews based on predefined rules
• Suggests improvements
• Flags potential issues

They iterate:

• Version 1 → feedback → Version 2 → feedback → … → Version N
• Eventually both “agree” → ready to ship

On paper, this sounds like a complete system.

But here’s the catch 👇

---

Why I Don’t Trust a Fully Autonomous AI Loop

Both the generator and reviewer are still operating within:

• learned patterns
• predefined rules
• limited business context

They lack:

• deep system understanding
• real-world tradeoff awareness
• domain-specific nuance

So while the output may pass “AI review,” it doesn’t guarantee:

• performance efficiency
• scalability readiness
• alignment with business constraints

---

My Approach: AI as an Assistant, Not the Owner

Here’s the workflow that works best for me:

Step 1: Generate Raw Material

I use AI to quickly produce:

• initial implementations
• boilerplate code
• multiple approaches

Step 2: Refine with AI

Then I use AI again to:

• optimize structure
• clean up logic
• identify obvious issues

Step 3: Human Final Pass (Most Important)

This is non-negotiable.

I manually:

• review logic deeply
• remove inefficiencies
• align with business use cases
• ensure long-term maintainability

---

The Interior Designer Analogy

Think of it like this:

You are a senior interior designer.

AI tools are your assistants:

• they prepare materials
• take measurements
• assemble components

But the final finish — the one clients notice — is yours.

Without that final touch, the output may be complete… but not refined.

---

Why the 90–10 Rule Matters

I strongly believe:

👉 90% of the work can be done by AI
👉 10% still requires human judgment

And that 10% is where the real engineering happens:

• performance tuning
• architectural decisions
• business logic alignment

---

Staying Connected to Code (And Why It Matters)

One hidden risk of over-relying on AI is losing touch with your own system.

If you depend entirely on AI:

• You may not fully understand what’s implemented
• Debugging becomes harder
• Decision-making becomes slower

With the 90–10 approach:

• You stay connected to the code
• You understand business flows deeply
• You can confidently make system decisions

So when someone asks:

“How does this feature work?”

You don’t need to go back to AI for answers.

---

Final Thoughts

AI coding is not a replacement — it’s a multiplier.

It accelerates development, reduces effort, and unlocks productivity.
But it doesn’t eliminate the need for engineering judgment.

The future isn’t:

AI vs Developers

It’s:

AI + Developers — with clear ownership

---

Over to You

I’m curious:

• Are you using a fully AI-driven workflow?
• Do you trust AI-generated code without manual review?
• Or do you follow a similar human-in-the-loop approach?

Let’s discuss 👇
Read more about how LLM thinks.