Nobody warns you about this: once you get significantly faster at your job, people start asking how.
It started with a workshop request. Then another. Over the past year, working as a software architect, I kept having the same conversation. An engineering lead pulls me aside after a demo and says:
"Okay, but how are you doing this? My team has Copilot licences. They have ChatGPT. They're not getting faster. Some of them are getting slower."
That last part isn't a hunch. It's been measured.
In 2025, METR published a study showing that developers using AI assistants without structured methodology were 19% *slower* than those working without AI at all. Not a rounding error — a measurable regression. Meanwhile, Anthropic's internal research found that the top 14% of AI-assisted developers reported productivity gains of two times or more. Early adopters report five to ten times (Anthropic Internal Study, 2025). heise Magazin confirmed the same pattern.
Same tool. Radically different outcomes. The variable isn't the AI. It's the human.
The Skill Nobody Teaches
My role has changed: I barely write code myself any more. Not out of carelessness — quite the opposite. AI tools with clean architectural guidance and clear constraints produce more focused code than I ever could alone. Stricter design patterns. No forgotten edge cases. No sloppiness at eleven in the evening. But only if you know how to direct them.
The industry calls this "vibe coding." Roughly as useful as calling surgery "vibe cutting." What most people actually practise is Vibe Prompting: type a wish into a chatbot, hope for good output, spend hours debugging the result.
Hand someone an orchestra and say: "Make music!" Every now and then a decent chord. Mostly noise.
The productive 14% work fundamentally differently. They think like solution architects before engaging the AI. Define the architecture. Set constraints. Establish quality gates. Specify what "done" looks like — before the first line of code is generated. Then they direct the AI within those guardrails: iteratively, methodically, with engineering discipline.
That's Vibe Engineering. Not prompting — orchestrating with architectural experience.
The uncomfortable truth: you have to learn it.
Like learning the violin, working effectively with AI requires deliberate practice. LLMs with poor architecture or vague instructions deliver the occasional usable result — but as soon as projects grow, it becomes critical. Bad input, bad output. Consistently. The gap between trained and untrained AI users is growing wider every month. It won't close on its own.
Workshops Don't Scale
After many workshop requests, I did the maths. Each session reached twenty people, at most. The methodology needed a few hours each time — not a set of tips, but a different way of thinking about software and processes that requires hands-on practice. At that rate: years to reach even a thousand developers.
And: Workshops don't measure anything. Participants leave energised, promise to apply everything — and three weeks later they're typing "build me a REST API" into GitHub Copilot. No architectural guidance, no validation by design. Decision-makers? No way to know whether the workshop changed anything. No evidence a manager could point to and say: This investment produced measurable results.
I knew what the solution had to look like. I'd spent many years building ML platforms. The pattern was obvious — a system that teaches the skill, measures the outcome, and delivers personalised coaching. A kind of Benny on-demand — a private coach, available any time. The problem: building a system like that normally requires a team and a six-month timeline.
But that was exactly what I wanted to prove — that Vibe Engineering works. So I built it over a weekend.
AI Teaching Humans to Use AI
Developers solve real engineering scenarios — not multiple-choice quizzes, but open-ended problems that mirror genuine design work. Each submission is evaluated by AI: across seven dimensions, against defined quality criteria, with personalised coaching. The platform uses AI to measure how well humans direct AI.
| Dimension | What it measures |
|---|---|
| Intent Clarity | Is the task unambiguously defined? |
| Contextual Grounding | Are relevant constraints and context provided? |
| Verification Strategy | How is the result validated? |
| Technical Leverage | Are the right tools being used? (MCPs, Skills, model selection) |
| Constraint Definition | Are boundaries and requirements clearly specified? |
| Decomposition Structure | Is the problem broken down into sensible steps? |
| Risk Resilience | Are failure scenarios and fallbacks considered? |
No black box that spits out a number. A radar chart across seven dimensions that changes shape as the developer improves. Trend lines showing progress over weeks. Coaching that explains not just what to improve, but how. No pre-built solutions to copy — Socratic reflection questions that force deeper thinking.
No pass/fail labels. Unlimited retries. Best score counts.
Most corporate e-learning works differently: watch the video, tick the checkbox, forget everything by Thursday. VibeSkills runs on a learning model: practise, get feedback, improve, practise again. Completion checkboxes tell you nothing. Score trajectories across seven dimensions tell you everything.
Nerd Talk: The Forty-Eight Hours
I want to be precise about what "built it over a weekend" means, because I've sat through enough startup pitches to know that phrase usually translates to: "Hacked together a demo that falls over if you look at it sideways."
The difference between Vibe Prompting and Vibe Engineering is exactly the difference that makes forty-eight hours possible. Vibe Prompting means typing "build me a training platform" into Claude Code and hoping the thing can go to production. Vibe Engineering means: before the first code is generated, thinking through the architecture, setting the constraints, establishing quality gates — and then forcing the AI to work within those guardrails.
Why this works is no secret:
Every word in the prompt changes the model's probability space. Vague instructions produce a random walk — after enough forks you're miles from what you wanted. Precise architectural specifications narrow that space so far that the AI can barely miss.
That's not magic. That's mathematics.
Here's what that looks like concretely: VibeSkills is a production-grade enterprise platform on AWS Frankfurt. One hundred percent EU-hosted. The architecture follows the Service Layer pattern (Fowler): 42 API routes, 19 services, 44 React components, 37 challenges. Testing Pyramid (Cohn) with 1,651 unit tests, 87 integration tests, 19 E2E specs, 108 evaluation regression tests. 86% branch coverage, 84% function coverage. A CI/CD pipeline that blocks any deployment that breaks a test.
Security? Not an afterthought — a constraint from line one. Every endpoint: authentication, rate limiting, input validation, sanitisation, audit logging. Not because I'm more disciplined than anyone else — because the AI implements non-negotiable requirements consistently when you define them as architectural constraints and include them in reviews. Across every route, every time. That's the point: The AI does exactly what's specified. No more, no less.
And when something does go wrong: unhandled exceptions are automatically analysed by AI — severity, root cause, affected component — and filed as a GitHub issue with full context. Another AI agent picks up the issue, creates a fix as a pull request, CI validates it, an AI agent checks it against the architectural guidelines, and the fix goes live. The same mechanism handles support: when a user reports a bug, the bot creates an issue, and AI fixes it autonomously. This isn't a future roadmap — it's running in production today. Self-healing by design.
Speed and quality aren't in tension when the architecture is right.
That's the core claim of Vibe Engineering. And VibeSkills is the proof — built with exactly the methodology it teaches.
Why Now
The EU AI Act entered into force in February 2025. Article 4 requires that employees working with AI systems have sufficient AI literacy. Enforcement by the Bundesnetzagentur begins in August 2026. Handing out Copilot licences doesn't count as evidence — Article 4 doesn't ask whether you provided tools, but whether your people are competent to use them.
McKinsey's latest research draws a sharp line: 80% of companies use AI as an efficiency tool — doing the same things slightly faster. The top 6%, the AI High Performers, use it as an innovation driver — new products, new business models that weren't possible before. When a product takes forty-eight hours instead of six months, you can run ten times more experiments. More experiments, more learning, better decisions. The compound effect is devastating for organisations still debating whether AI is "ready for production."
The arithmetic: a developer at 2x produces the output of two. A fifty-person team at 2x equals a hundred-person team. The difference between the untrained team getting slower and the trained team doubling output — that determines which products ship and which don't.
The gap between the productive few and the rest isn't talent. It's methodology. And methodology can be taught. I know, because I built the system that teaches it. In forty-eight hours.
If your team has Copilot licences and productivity still isn't improving — the tool isn't the problem.
Watch the platform in action:
I welcome feedback — reach out directly or through the platform.
Top comments (0)