They call it vibe coding. I have been doing it 6-8 hours a day for about a year. Let me try to describe what really works, what does not, and how my craft has changed.
A premise: I am not an AI evangelist. I am a developer who, at 32, found himself with a tool that flips how I write code, and that — as happens every so often in this industry — demands to be understood rather than loved or hated.
What I mean by vibe coding
It is not "I press Tab and Copilot completes a function". That is autocomplete on steroids, it has been around since 2021, it is an IDE feature.
Vibe coding is: you open a conversational session with a strong model (Claude Sonnet or Opus in our stacks), you give it access to the project filesystem and to the shell, and you hand it tasks that until yesterday required a junior intern. "Add a /api/users/me endpoint with cursor-based pagination", "Migrate this library from Pinia to Zustand", "Find out why the 'cart total' test fails only in CI". And it opens files, writes, runs, retries, and asks you only when it has to.
The practical difference: with autocomplete, you think and the AI fills in. With vibe coding, the AI thinks inside its loop and you do code review. It is a role reversal not everyone accepts. Those who do work 2-4 times faster.
What we actually write this way
In the last 90 days we have shipped MVPs for early-stage startups, internal dashboards for SMEs, custom MCP plug-ins. As a share, code written with vibe coding is around 70% of the total. The remaining 30% is the delicate stuff — domain models, payment logic, security controls — handled by hand by a senior, perhaps asking the AI for a second opinion but never delegating the writing.
What works well:
- ◆Bootstrapping an MVP from scratch. A "Next.js + Supabase + Stripe + Resend" pipeline: Claude Code stands it up in 4 hours, deploys it in 5.
- ◆Mechanical refactors. "Move every fetch into a central client", "drop styled-components, switch to standard CSS-in-JS". Large-scale pattern matching.
- ◆Integration tests. It writes tests that mimic real API traffic, including edge cases we humans would be too lazy to cover.
- ◆Documentation that stays in sync with the code. Tell it once and every PR includes it in the to-do.
Where it still struggles:
- ◆Performance optimisation that requires understanding the hardware. It picks defensive library choices, missing the clever wins.
- ◆System architecture beyond a certain complexity threshold. It cannot see the whole.
- ◆UI when the design is not already encoded in design tokens. It improvises, and the improvisations are mediocre.
The setup we use
A clean Git repo. An AGENTS.md (or CLAUDE.md) at the root describing conventions — naming, folder structure, preferred libraries, test patterns. The agent reads it at the start of every session and adapts.
We launch Claude Code in the terminal, inside the project folder. No complicated IDE integration. We watch the output together, we step in when it strays. A session typically lasts 30-90 minutes, at the end there is a PR ready.
Code review is the skill that has grown most in the team. Reading code fast, judging whether the structure is right, trusting your eye: those are the muscles that get trained today. Writing single perfect functions is an increasingly less distinguishing skill.
How much we are saving
We measured it across 4 comparable projects over the last year: same stack, same kind of feature, equivalent size.
- ◆Without vibe coding: 22 days on average to first deploy.
- ◆With vibe coding: 9 days on average.
The factor is not uniform. On "initial bootstrap" (auth, DB schema, first endpoints) we go 4-5x faster. On "final polish" (pixel-perfect, business edge cases) we laugh at our old figures because we are back to human pace there. Averaged out, it is about 2.4x.
When we do not use it
When we work inside a client's codebase with heavy compliance (healthcare, banking) and we cannot use external models, vibe coding disappears. We go back to Cursor with Tab completion and nothing else. The hit on time is large — easily 3x slower — but it is the reality with certain clients.
When the task is almost all thinking and almost no writing — refactoring a legacy system whose business rules are not in the code but in the CTO's head — the AI has no leverage, and the only path is a call with the CTO on speakerphone.
What clients think of it
Most of them care little about how we write. They want quality, timelines, prices. Vibe coding helps with timelines without degrading quality (in fact it raises it, because we have more time for the delicate parts). On price, some clients ask for a discount because "the AI is helping you". We answer that we price for delivered value, not keyboard hours. The conversation ends in two lines.
What I think
My craft has changed more in 12 months than it had in the previous 10 years. It is destabilising and beautiful. Those entering this work now have a brutal advantage: they do not have the instinct to write by hand things a model writes better. Those who have been in it for 15 years have another advantage: they sense when code is rotten even when it looks pretty, and models still do not sense that.
Balanced like this, it is the best moment to write software that I can remember.