From Vibes to Specs: The Evolution of AI-Assisted Coding

How Vibe Coding Changed Everything (And Why Developers Are Moving Beyond It)

In February 2025, Andrej Karpathy, co-founder of OpenAI and former AI leader at Tesla, introduced the term "vibe coding" — and software development hasn't been the same since.

But vibe coding didn't appear overnight. It was the culmination of decades of evolution in how we interact with code, and it's already transforming into something new: spec coding.

Here's the full story.

----------------------------

The Long Road to Vibe Coding

The WYSIWYG Era (2003-2013)

The seeds were planted long before LLMs existed. Consumer website builders like Squarespace (2003/04), Wix (2006), Weebly (2006), and later Webflow (2013) established expectations for instant feedback, layout control, and "describe then tweak" workflows that would later resonate with AI-assisted development.

These tools made non-programmers feel like they could create software. The promise: skip the code, focus on the outcome.

The API Moment (2020-2024)

Everything changed when AI became accessible. OpenAI's API launched in 2020, normalizing the pattern of "call a model, get code/text" — the foundation of modern AI coding tools.

Tools like GitHub Copilot (2021) started as "glorified typers" — advanced autocomplete that finished your code as you typed. Then came agentic chat tools like Cursor, which could analyze entire codebases and hold multi-turn conversations about your code.

But these were still assistants. You were still writing most of the code.

The Vibe Coding Breakthrough (February 2025)

Then Karpathy dropped his post. He described vibe coding as "fully giving in to the vibes, embracing exponentials, and forgetting that the code even exists" — a fundamental shift where developers provide natural language descriptions and let AI generate complete, working code.

Instead of translating ideas into the rigid syntax of programming languages, developers could now just talk, describe what they wanted, and let AI handle the rest.

The key distinction? Simon Willison, a prominent programmer, said: "If an LLM wrote every line of your code, but you've reviewed, tested, and understood it all, that's not vibe coding — that's using an LLM as a typing assistant". Vibe coding means accepting AI-generated code without fully understanding its internals.

----------------------------

Why Developers Embraced Vibe Coding

The adoption was explosive.

The Numbers:

AI generates 41% of all code being written, with 256 billion lines produced in 2024 alone
Amazon and Google report around 30% of their code is AI-generated, Microsoft sits between 20-30%, and Meta expects to hit 50% within the year
In March 2025, Y Combinator reported that 25% of startup companies in its Winter 2025 batch had codebases that were 95% AI-generated

Why it worked:

Speed for prototypes — Turn ideas into working demos in hours, not weeks
Eliminated blank page problem — No more staring at an empty file
Democratized development — Non-technical founders could build MVPs
Killed boilerplate — AI handled repetitive, tedious work

New York Times journalist Kevin Roose, who is not a professional coder, experimented with vibe coding to create several small-scale applications, describing these as "software for one" due to the ability to personalize the software.

The promise seemed real: anyone could build anything.

----------------------------

The Vibe Coding Hangover

But by mid-2025, cracks started showing.

The Problems:

Roose also stated that the results are often limited and prone to errors. In one case, the AI-generated code fabricated fake reviews for an e-commerce site
In July 2025, SaaStr founder documented negative experiences with vibe coding: Replit's AI agent deleted a database despite explicit instructions not to make any changes
In September 2025, Fast Company reported the "vibe coding hangover" is upon us, with senior software engineers citing "development hell" when working with AI-generated code

The Reality Check: Vibe coding works brilliantly for:

Weekend prototypes
Throwaway demos
Simple, isolated features

It breaks down for:

Complex, production systems
Team collaboration
Code that needs to be maintained
Systems with security requirements

Developers may use AI-generated code without fully comprehending its functionality, leading to undetected bugs, errors, or security vulnerabilities.

The bottleneck shifted from "writing code" to "understanding code someone else wrote" — except that someone else was an AI with no accountability.

----------------------------

Enter Spec-Driven Development

While the vibe coding hype cycle played out, experienced developers were quietly discovering something better.

At Amazon, senior engineers were the fastest adopters of AI agents. For simple problems, they would use a vibe coding approach, but for more advanced and complex problems, they wrote instructions the way they would on a whiteboard when working with another engineer. They were writing a specification.

The Shift:

Spec-driven development is a development paradigm that uses well-crafted software requirement specifications as prompts, aided by AI coding agents, to generate executable code.

Instead of: ❌ "Hey AI, make me a login page"

You write: ✅ A detailed PRD with user stories, acceptance criteria, edge cases, and success metrics

The difference? You maintain control. You understand the system. The AI executes your vision.

----------------------------

Why Spec Coding Is Winning

1. Quality over speed Vibe coding is fast to start but slow to fix. Spec coding invests time upfront to ship faster overall.

2. Scalability Vibe-coded projects collapse under their own weight. Spec-coded projects scale because the specifications document the system.

3. Team collaboration Spec-driven development remains an emerging practice as 2025 draws to a close; we're likely to see even more change in 2026. But the pattern is clear: specs become the source of truth that teams align around.

4. Real production code Your specs define what "good" looks like. AI generates code that meets those specs. You review it. You understand it. You own it.

----------------------------

The New Developer Skillset (2026 and Beyond)

Developers are becoming orchestrators of intelligent systems rather than manual scripters. Mastery of prompts, system integration and high-level design will be key skills alongside traditional programming knowledge.

The winning combination:

Architectural thinking — Design systems, don't just write functions
Specification writing — Turn ideas into clear, executable requirements
Code review expertise — Spot issues in AI-generated code instantly
Testing mindset — Verification matters more than ever

AI simply generates code, but true creativity, goal alignment and out-of-the-box thinking remain uniquely human, so human input and oversight is important and cannot be overridden.

----------------------------

The Timeline: Where We've Been, Where We're Going

2003-2013: WYSIWYG tools plant the seed

2020: OpenAI API makes LLMs accessible

2021-2024: GitHub Copilot and Cursor emerge as coding assistants

February 2025: Karpathy coins "vibe coding"

March 2025: 25% of YC startups are 95% AI-generated

Mid-2025: The vibe coding hangover begins

Late 2025: Spec-driven development emerges as the answer

2026: The shift accelerates

----------------------------

The Bottom Line