# Vibe Coding Since 2021

Remember when we used to actually *write* code? How cute. Back in August, 2021, while most developers were still busy arguing about tabs versus spaces or which JavaScript framework would die next, OpenAI quietly dropped a steamer called Codex. At this point, only the most die-hard AI enthusiasts were paying attention. Fast forward to June 2022: GitHub Copilot becomes generally available, and over the next year or so, developers start to watch AI complete their for-loops with a mix of awe and existential dread. "It's just spicy autocomplete," we reassured ourselves, clinging to our Stack Overflow reputation points like life rafts. Plot twist: We weren't witnessing a better code suggestion tool. We were watching the birth of an entirely new programming paradigm that would eventually be named by Andrej Karpathy's viral [tweet](https://x.com/karpathy/status/1886192184808149383?s=61) in early 2025: "vibe coding." Whether it's the new models, or the notion of "vibe coding" itself, the last two months have seen a real vibe-shift in terms of the mainstream adoption of AI coding tools. Here's how we went from copy-pasta coding to zero-shot full-stack apps in just a few short years. ## Phase 1: Completion (2021-2022) > *""it's a really really good pattern completion system that happens to work on patterns in code, it's like the world's best "yes, and" improv actor whose domain happens to be code rather than improv."*

GitHub's Copilot technical preview announcement video showcased what seemed like magic at the time.

The first era of AI coding tools was primarily about completions. As an early GPT-3 Beta user, I was fortunate to receive my invite on **August 19th, 2021**—a date forever etched in my memory as the beginning of my AI journey. Those early days in the OpenAI Playground were simultaneously thrilling and frustrating—seeing a large language model complete my thoughts felt like science fiction made real, but the results were wildly inconsistent. When Codex (the code-specialized version of GPT-3) arrived, it represented a quantum leap. Suddenly, typing a comment like `// create a function that sorts an array of objects by date` would produce a working implementation. For those of us in the Beta, it felt like we'd been given a glimpse of the future—one where programming would fundamentally change. A pivotal moment happened on **September 10th, 2021**, when I created my first "vibe coding" demo. Though we didn't call it that yet, this experiment showed that AI could generate coherent code from abstract intentions rather than explicit instructions. Looking back, this was the prototype for what would later evolve into full-fledged vibe coding.

OpenAI's Codex demo showcased capabilities that seemed magical at the time—watching it generate entire functions from natural language prompts felt revolutionary.

By the time GitHub Copilot officially launched as a technical preview in June 2022, the early experiments were becoming a commercial reality. Developers across the industry began experiencing what only beta testers had seen before. Despite the excitement, the limitations of this early era were evident: - AI would complete code based on the local context only, which itself was only 1024 tokens - Codex and early Copilot often "hallucinated" APIs that didn't exist - The models struggled with complex logic and architectural understanding - You needed to be exceedingly explicit in your comments and prompts My workflow in those days followed a predictable pattern: Write a fuzzy comment describing a function (possibly Googled), watch the AI generate a first pass, copy paste the code into my IDE, manually edit the 40-70% that was correct, and then repeat until it ran. People talk about their "aha" moment with AI coding, and this was around the time when I had mine. ## Phase 2: Conversation (2022-2023) ### *"ChatGPT changed everything"* When ChatGPT launched in **November 2022**, it reshaped our relationship with AI almost overnight. This wasn't just an iterative improvement—it was a paradigm shift. Suddenly, we weren't just receiving completions—we were having conversations. This marked the transition to a much more fluid, dialogue-based approach to development. GitHub Copilot Chat, Claude, and other tools rapidly evolved to incorporate this conversational paradigm into the development environment. The change was profound: - We could ask questions in natural language - We could request explanations of complex code - We could describe what we wanted to build in plain English - We could iterate through ideas by conversing with the AI My workflow transformed: Instead of writing code first, I started describing what I wanted to build, letting the AI generate a first pass, then refining through conversational feedback. It still felt like I was directing the process, but the AI became more of a peer than a tool. ## Phase 3: Vibes (2023-2024) ### *"Just transmit the vibe and let the AI manifest it"* Somewhere in late 2023, a subtle but profound shift occurred. The relationship between developer and AI inverted. Instead of the developer writing code with AI assistance, developers began to transmit "vibes"—abstract intentions, design philosophies, and desired outcomes—and let the AI handle the implementation details. This was the true birth of "vibe coding" as we now know it. The hallmarks of this era: - Sparse, high-level prompts replaced detailed specifications - AI became responsible for architectural decisions - Developers focused on evaluating and refining rather than writing - The emphasis shifted from syntax to semantics I remember the first time I truly "vibe coded." I had been struggling with a complex state management problem for hours. Frustrated, I finally just wrote: "I need a clean way to manage global app state that doesn't cause re-render issues. It should be intuitive for other devs and performant." The AI gave me a complete, elegant solution that I wouldn't have come up with myself—not because I couldn't, but because I was thinking too narrowly. It had internalized patterns from thousands of similar problems and synthesized an approach that transcended my own mental models. That moment changed me as a developer. The release of **Claude 3.5 Sonnet integrated with Cursor in October 2024** marked another inflection point. The combination of Cursor's codebase understanding and Claude's reasoning capabilities created a tool that could truly understand project context and architectural patterns, not just isolated code snippets. ## Phase 4: Symbiosis (2024-Present) ### *"We're forming a new kind of collaborative intelligence"* The current phase, which began emerging in early 2024 but truly crystallized with the release of **Claude 3.7 with Cursor Agent Mode and MCP in February 2025**, is what I call "symbiosis." This technological leap brought agent-based autonomy and Model Context Protocol to development, allowing the AI to maintain complex mental models of entire codebases and development contexts. The boundaries between human and AI contributions have blurred to the point where the output is a true fusion—neither fully human nor fully AI, but a new kind of collaborative intelligence. In vibe coding symbiosis: - Developers transmit abstract intentions and creative direction - AIs handle implementation details, patterns, and optimizations - Development becomes iteratively conversational and intuitive - The distinction between "my code" and "AI code" dissolves My current workflow is almost unrecognizable compared to 2021: 1. I communicate a high-level intention or problem statement 2. The AI generates a comprehensive solution 3. I provide feedback at the conceptual level ("make it more maintainable" or "prioritize readability over performance here") 4. The AI refines based on my abstract direction 5. We iterate until the solution feels right ## Looking Ahead: 2025 and Beyond As we look to the future, it's clear that vibe coding will continue to evolve. The symbiosis will likely deepen as models become more capable and our ability to communicate with them improves. I expect we'll see: - Multimodal inputs becoming standard (voice, sketches, gestures) - AI taking on more autonomous maintenance and refactoring - Emergence of specialized AI collaborators for different domains - New programming languages designed specifically for human-AI collaboration - A renaissance of ambitious software projects previously considered too complex The most profound shift, though, has already happened: We've redefined what it means to be a developer. Writing code is no longer the primary skill. Thinking clearly, communicating intentions, and having a vision for what should exist—these are now the core competencies. The journey from "fancy autocomplete" to collaborative intelligence has been shorter than any of us imagined. And much like the journey itself, this blog post was a collaboration—my ideas and experiences, shaped and enhanced by the same AI technology I'm reflecting on. I'm not sure where the line is between my thoughts and its contributions anymore. And I'm not sure it matters.