Service 03
Vibe coding
Having AI build software from plain-language descriptions rather than written code. It works remarkably well for the first prototype and remarkably badly for whatever comes after. I help you find that line and cross it deliberately.
Why the first weekend goes so well
Vibe coding produces something that works within hours. That part is real, not marketing. The trouble starts at version two: nobody understands why the code does what it does, there are no tests, and the first time something breaks nobody knows where to look.
I see the pattern regularly now. Someone builds a tool over a weekend that solves an internal problem, the team starts using it, and six months later business-critical software is running that nobody dares touch. That is not an argument against vibe coding. It is an argument for knowing when to switch modes.
What the service covers
Depending on where you are: building alongside you, or cleaning up what already exists.
-
Building fast where speed is the point
Internal tools, prototypes, one-off scripts and dashboards. Exactly the work vibe coding is best at, because the cost of a mistake is low.
-
The move to production
If the prototype works and you want to build on it, something has to happen: tests, error handling, readable structure, documentation. Usually that is less work than starting over.
-
Reviewing what you already built
If you built something yourself and are unsure whether you can build on it, I look at it and tell you honestly whether it will hold. Sometimes the answer is to throw it away and start again.
-
Teaching your team where the line sits
The most valuable outcome is usually not the code but the judgement: when vibe coding is fine and when you need an actual developer.
How it runs
- 01
Decide what it needs to become
A throwaway prototype and a system your invoicing will run on need completely different treatment. We make that choice up front, not halfway.
- 02
Build with AI, review by hand
The speed of AI, with judgement applied to what comes out. Particularly wherever money, customer data or external integrations are involved.
- 03
Hand over or shut down
If it stays running, I make sure someone understands it. If it was a prototype, we throw it away, and that is fine.
What you end up with
Working software in days rather than months, without being locked in six months later to something nobody can maintain.
- Internal tools finished in days rather than in a sprint plan
- Clarity on what is and is not production-grade
- Existing AI-written code that is tested and documented
- A team that can judge for itself when the risk gets too high
What it costs
What an engagement costs depends on its scope and on how many systems are involved. I work with a fixed price per phase rather than open-ended billing: you know what a phase costs before it starts, and you decide after each one whether to continue.
I only quote a figure once I know what we are talking about. That takes a half-hour conversation, at no charge.
Whether this fits
This works well when
- you have an internal problem not worth six months of development
- something has already been built with AI and you want to know if it holds
- you need speed but do not want to fly blind on the result
- it is clear who will actually use the tool
This fits less well when
- the software is something you sell to customers
- sensitive personal data or payments flow through the system
- you expect AI to take over the maintenance too
Questions about vibe coding
What exactly is vibe coding?
Having an AI model write software from plain-language descriptions, steering on the result rather than on the code itself. The term comes from Andrej Karpathy, who described in early 2025 how he programmed by letting the AI write and barely looking at the code himself. For small, bounded things it works strikingly well.
Can this replace my developer?
For internal tools, scripts and prototypes, often yes. For software your business depends on, no. The difference is not whether the code works but what happens when it breaks: then you need someone who understands what is there. Anyone claiming AI fully replaces developers has not yet lived through a production incident.
Is AI-written code secure?
Not automatically. AI models regularly produce code with known weaknesses: insufficient input validation, keys committed into the source, missing permission checks. For an internal tool with no sensitive data that is an acceptable risk. As soon as customer data or payments are involved, someone who knows what they are looking at needs to review it.
Let's discuss how I can help your business grow
A short call is enough to tell whether there is a good fit. No sales pitch, just an honest read on what is possible.