AI

How AI coding Evolved from Vibes to Specs

created by ChatGPT-5

Vlado Balko

Oct 23, 2025

Tracing the Evolution of AI-Assisted Development: From Vibe Coding to Structured Workflows

This post is part of my ongoing series about vibe coding — tracing how AI-assisted development evolved from early experiments to structured, spec-driven workflows.

Previous posts:

When AI Crafts Games: A Peek into Building a JS Game with ChatGPT
Exploring the Hype Around Claude 3.5 Opus by Anthropic
Press enter or click to view image in full size
Vibe coding timeline

When the Code Started Talking (End of 2022)

It all began at the end of 2022, when ChatGPT-3.5 arrived and suddenly the code started to talk back. Until then, coding had always been a solo act. You wrote, compiled, debugged, and maybe cursed under your breath. But with ChatGPT-3.5, something shifted. You could ask for a snippet and get a working solution. I probably started with some SAP UI5 component or Groovy script for SAP SCI. It wasn’t perfect, but it was faster than anything I’d known before.

Memory was short — 8k tokens — but we didn’t know any better, so we adapted. You learned to live with amnesia, saving context by committing code obsessively, terrified of losing that one magical generation. Yet even with such constraints, it felt like a revolution. ChatGPT-3.5 wasn’t just an assistant — it was a curious, clumsy coding buddy who occasionally hallucinated but almost always surprised you.

The Vibe Coding Era (GPT-4 → Claude 2/3)

Then came GPT-4 and Claude 2, and suddenly the vibe had rhythm. Context windows grew, hallucinations dropped, and intent recognition improved. It was no longer about brute-force prompting but about co-creation. For many, this period defined what became known as vibe coding — that intuitive, conversational style of programming where you could just say, “I have an idea — can you make it for me?”

At first, it felt like advanced autocomplete — something between GitHub Copilot and a creative brainstorming partner.
Tools like Copilot X and Cursor IDE made it even smoother.
By mid-2023, you could chat through an entire feature or refactor with minimal manual typing.
The community exploded with examples — YouTube tutorials, Medium posts, and wild X threads where people built full apps by “vibing” with AI.

We went from hyped prototypes to structured experiments in context and creativity.

The Agent Awakening (Mid-2024)

The real game-changer came with the agentic tools. Cursor’s agent mode hit like a flash from a clear sky: it didn’t just write code — it ran it, debugged it, and fixed its own mistakes. The first time I saw it execute and rerun without being told, I just stared at the screen for a few seconds in awe. The feeling was somewhere between excitement and slight dread — maybe 80% thrilled, 20% worried about what could go wrong.

Then came Claude Code and Gemini CLI, followed by GPT Codex and smaller experiments like Kiro.

Suddenly, I wasn’t just coding with AI — I was managing a team of digital juniors.
You had to learn to shepherd them, like a flock of eager but unpredictable sheep.
Let them run in YOLO mode and they’d break things.
Guide them carefully, and they’d deliver elegant, usable results.

Context Engineering and the PRD Phase (Late 2024 — Early 2025)

As agents improved, a new problem emerged: context loss. Your once-expert AI partner could wake up one morning lobotomized — forgetting your project, your style, your process. The fix wasn’t more prompting but better structure. That’s when I started writing PRDs (Product Requirement Documents) for AI.

I adopted GitHub SpecKit’s structure: Spec → Plan → Tasks. It made perfect sense — teach the model what it’s building, then how to build it. The discipline of writing structured specs transformed my workflow. I learned that good folder organization and clean code weren’t just for humans; they helped AIs handle context compression better. Ironically, vibe coding led me straight into structured project thinking.

Context persistence tools — SpecKit, OpenSpec, and others — offered huge improvements, but none solved it completely.
Even today, context degradation remains the biggest challenge in long-form AI development.

Post-Vibe Coding and the Context-Persistent Era (Mid-2025 → Present)

By 2025, the vibe evolved. The tools matured, and so did we. Post-vibe coding arrived — not as the death of creativity, but as its discipline. GitHub SpecKit, OpenSpec, BMAD, and PRP brought persistent specs and reasoning chains. Now, AI could maintain continuity between sessions, remember design decisions, and act as a structured collaborator.

Preparation time skyrocketed — you spent hours crafting a spec — but the payoff was undeniable.
The AI delivered cleaner code, uncovered hidden edge cases, and forced you to think like an architect.
It’s not just faster coding anymore; it’s smarter building.

Reflection: The Shepherd of the Bots

Looking back, I moved from tinkerer to collaborator to what I jokingly call a shepherd of bots. My job is no longer writing code line by line — it’s orchestrating context, setting direction, and guiding a team of tireless digital apprentices.

If there’s one piece of advice for anyone entering AI-assisted development, it’s simple: don’t be afraid.

The worst that can happen is you lose some code (so commit often).
Experiment, iterate, and treat the AI like a partner that learns as you do.

Documentation-first thinking isn’t new — it just trickled down from large enterprise systems into individual workflows. Thanks to AI, even small prototypes now deserve structured specs, architecture planning, and proper documentation. And with that shift, we’ve become one-person development teams: architects, designers, testers, and developers all in one.

The revolution isn’t about machines replacing us — it’s about learning to collaborate at scale with them. The vibe lives on, just wearing a suit now.
Press enter or click to view image in full size
Still coding, still vibing — but now with specs.