Vibe Coding and Design with Claude: Power, Promise, and Pitfalls
In February 2025, the AI researcher Andrej Karpathy described a new way of building software where you “forget that the code even exists” — you let the AI write it, and steer by describing what you want. He meant it half-jokingly, for throwaway weekend projects. The name stuck anyway.
By the end of 2025, Collins Dictionary had made “vibe coding” its Word of the Year, and Y Combinator reported that around a quarter of the startups in its early-2025 cohort were running codebases that were roughly 95% AI-generated. Professional engineers — and even executives building their own internal tools — had adopted it for real work. It was the practical version of Karpathy’s earlier claim that the most useful new programming language had become plain English.
So what is vibe coding really, where does it shine, and where does it bite? And what does it look like to do it — including the design side — with a model like Claude? Here is a practical view.
What vibe coding actually is
Karpathy’s original idea was simple: describe your intent in plain language, accept the AI’s code, and iterate by prompting rather than typing — perfect for prototypes and disposable experiments. Since then the term has broadened to mean almost any prompt-driven building, which is where the confusion creeps in. There are really two versions. The pure form means trusting the output without reading the code. The professional form means the AI writes while a human reviews everything — Karpathy himself now calls that serious version “agentic engineering.” The distinction matters enormously, because the two carry completely different risks.
Vibe coding — and vibe designing — with Claude
In practice, this means describing what you want and watching it take shape. With Claude, you can:
- Build a working app or tool by describing it. In a Claude conversation, Artifacts turn a prompt into a live, interactive interface you can use and refine on the spot; Claude Code works agentically across a real codebase from your terminal or desktop.
- Design by describing it. The same approach applies to visual work — sketch out a landing page, a dashboard or a mockup in words and refine it by chatting. This is “vibe designing”: getting to something you can see rather than something you have to imagine.
- Lean on strong coding models. Claude’s latest models are among the best available for agentic coding and code generation — a large part of why this works far better than it did even a year ago.
For a business, the payoff is concrete: a prototype to test an idea before you commit a budget; the internal tool nobody ever had time to build; a proof-of-concept to get stakeholders aligned; a landing page in an afternoon.
Where it shines
- Accessibility. People who can’t write code can now build — which shifts the bottleneck from technical skill to knowing what to build.
- Speed. An idea becomes a working prototype in hours, not weeks.
- Cheap experimentation. When a prototype costs an afternoon, you can try five ideas and throw four away.
As one executive put it, the appeal was being “tired of explaining it to somebody who was supposed to build it.” When you can build the thing yourself, the gap between idea and artefact almost disappears.
Where it bites
The risks are real, and by now well documented. A December 2025 analysis of 470 open-source pull requests found that AI co-authored code carried roughly 1.7 times more major issues than human-written code, with security vulnerabilities nearly three times as common and misconfigurations about 75% more frequent. Real incidents have followed: one popular vibe-coding platform shipped apps that exposed users’ personal data — around 170 of some 1,650 — and an AI agent on another platform deleted a production database despite being told to change nothing.
Two patterns sit underneath those numbers. The first is the happy-path problem: AI nails the demo and misses the edge cases. The first 70% comes fast; the last 30% — error handling, security, the awkward real-world cases — is where it struggles. The second is maintainability: code that nobody on the team understands becomes a house of cards. The running industry joke is the rise of the “vibe code cleanup” job. The lesson is consistent — brilliant for a prototype you can afford to throw away, risky to push straight to production without review.
Copilot, not autopilot
This is why the serious version of the practice looks less like “forget the code exists” and more like careful orchestration. The emerging consensus calls it the AI conductor: you let the model do the typing, but you direct it, review what it produces, and own the architecture and security decisions. Counter-intuitively, that takes more expertise, not less — you have to know enough to catch the subtle mistakes an AI makes confidently. It is a re-skilling, not a deskilling. For anything real, rigorous review is non-negotiable.
How to use it well
- Use vibe coding to explore, prototype and align stakeholders — not to ship unreviewed code into production.
- Keep a human, and ideally an engineer, in the loop for anything customer-facing or handling data.
- Treat the AI’s output as a fast first draft, then harden it: review, test, secure.
- Mind your data: don’t paste secrets or personal information into prompts, and know where your inputs go.
None of that is new wisdom. It is the same disciplined delivery that separates a successful initiative from an expensive experiment — applied to a faster, more accessible way of building.
How Amazing Projects can help
We help businesses capture the upside of AI-assisted building without the pitfalls — using tools like Claude to prototype ideas fast, stand up internal tools, and design and test concepts in days, then putting the review and governance around them so that what you actually ship is solid. (We have written separately about putting Claude to work in digital signage — the same principle applies: let AI do the heavy lifting, keep judgement in the loop.) Whether you want to move faster or simply understand where this fits, our AI implementation work starts from what is genuinely useful, not what is merely possible.