Vibe Coding: The Prototype Is the Easy Bit

by | Nov 16, 2026

16 November 2026 · by Marcie D. Terman

Every business has a list

We opened our seminar, Vibe Coding: What business problem is driving you nuts?, by asking attendees one question: what’s the task in your business that keeps driving you crazy? The answers arrived within seconds from multiple people drawn into the meeting. We’ll touch on them later in this post

Start with the right problem

Not every problem is a vibe coding problem. As DATAFORT co-founder Gavin Smith put it, “Vibe coding software will happily have a go at anything, but that’s not necessarily going to mean you end up with a useful result.” A good vibe coding project needs to pass three tests:

  • It’s bounded. You can describe what “done” looks like in a paragraph. If you need a 40-page requirements document, you’re not vibing; you’re running a programme.
  • You can check it’s right. Someone in the room knows what good looks like and will spot a wrong answer fast. If nobody can tell, “that’s not vibe coding, that’s a hope and a prayer”.
  • It has a low blast radius. If it falls over on a Tuesday, nobody loses money, nobody gets fined and nobody has to phone the regulator.

The ideal candidate for vibing is the spreadsheet with 17 tabs and a macro only Dave understands: real pain, done by hand every week, with a grateful user who will happily test it for you. Poor first candidates are core transaction systems, anything touching payments, anything security-critical and anything regulated, where “mostly works” can end up an expensive error due to lost money or regulatory fines.

Two more rules of thumb from the session:

  • Solve the problem, not the hype. Using AI to build a tool doesn’t mean the finished tool needs AI running inside it. Most business problems are solved by a form, a database, a few rules and a report: boring, reliable and cheap to run.
  • Learn to drive before you hand the keys to the robot. Autonomous AI agents (AI that takes multi-step actions on its own) are powerful, but they are the right answer for a small set of genuinely open-ended problems. They are not a first project.

What a well-run build looks like

To show the process honestly, Austin Cooper, DATAFORT’s DevOps and security manager, took a real problem from DATAFORT’s own operations manual. We use a licensed service, Calendly, for booking calls and client meetings from our WordPress website. Could vibe coding replace it with something we own that perfectly meets the company’s needs?

It took about two hours. What made it work wasn’t the AI writing code; it was the structure around it:

  1. A brief before any code. The AI studied how Calendly was used across our site, how it integrated with Google Calendar and how it allowed staff members to individually set their own availability, then produced a 12-stage plan covering build, testing and security. A person reviewed that plan before anything was built.
  2. Hard boundaries. The AI was not allowed to deploy to the live site, use real customer data or communicate outside its own test environment. Within those limits it worked on its own, coming back only when it was genuinely stuck.
  3. Testing that tries to break things. It sent 40 simultaneous requests for a single one-to-one slot: one was booked, 39 were turned away. It then switched off its own double-booking protection to prove the test would catch the failure. It did: all 18 requests got the same slot.
  4. A second opinion. A separate AI agent, given a fresh copy of the code, reviewed security independently and tightened up the Google Calendar connection.
  5. A human acting as a customer. Trying it as a visitor revealed a time zone dropdown with around 400 options, unusable on a phone. The fix went back to the AI with a screenshot of what good looked like.

The finished booking system is now live on datafort.com, almost identical in look, feel and functionality to what it replaced, and it saves the company a monthly subscription that increases as the number of staff grows.

Where prototypes stall: integration, security and reliability

A working prototype is the start, not the finish. The moment colleagues depend on it, it holds real data, or it needs supporting by someone who wasn’t there when it was built, it has to graduate to proper engineering. That’s where most DIY projects get stuck, and it’s where our experience counts.

Security starts at the first prompt

An AI that thinks it’s writing a throwaway script will write a throwaway script: no login, passwords or settings baked straight into the code where anyone can see them, and everything wide open. Tell it from the outset where the software will run, who will use it, what it connects to and where the data lives (especially if customer records need to stay in the UK or EU), and it makes very different choices. One client Gavin spoke to had built a polished prototype with a popular vibing tool, only to find there was absolutely no security layer. They had to start the engineering again.

Rules that matter belong in code, not in polite requests

“Never delete records” in a prompt is an instruction the AI weighs up, not a law it obeys. Given a clever enough input or a long enough conversation, it can talk itself round. If a rule genuinely matters, enforce it in the system: read-only access to the database, or a hard £500 limit on payments.

To judge how much control you need, ask three questions:

  • Is the input trusted? Your own team typing into an internal tool is one thing. A customer email, an uploaded document or a web page is not.
  • Is the action reversible? A draft someone approves can be deleted. Money sent, records erased and emails to clients can’t be taken back.
  • How critical is the process? Tidying meeting notes is not the same as touching the ledger, customer data or anything a regulator cares about.

Pass all three and guidance in the prompt is probably fine. Fail one and you need controls built into the system you are vibe coding. Fail two or more and you also need a person to sign off before anything happens.

Reliability is everything around the code

Gavin’s advice is to treat AI-generated code like work from “a very fast, very confident junior”: often good, occasionally spectacularly wrong, and never in doubt. In practice that means:

  • Keeping a save-history (version control) and reviewing what it produces, so you can roll back to an earlier version if a new update breaks something.
  • Giving the software the least access it needs, through its own account, never an administrator’s login.
  • Watching what the AI “fixes”: when something won’t work, models like to switch off security to make it work.
  • Checking every external code library or add-on it installs. Models occasionally invent ones that don’t exist, and attackers register those names.
  • Keeping passwords and security keys out of prompts entirely.
  • Logging what the system saw, decided and did. “The AI did it” is not an audit trail.

And then there’s integration: connecting a new tool to the systems you already run, without opening new gaps. That’s rarely a vibe coding job on its own.

Your questions, answered

Delegates brought real problems in the chat and by email beforehand. Here are two examples that show the thinking in action:

  • Sarah’s problem: “I want an alert when an AI-drafted quote is ready to check.” No AI needed for this part. A simple checker with read-only access confirms the quote is complete and raises a flag, and the existing tool waits for it. Keeping it separate keeps AI out of the decision.
  • “Our office runs on one person’s spreadsheet.” A textbook first project. Use a dedicated AI coding tool (like Cursor or Replit) rather than a standard chat window, show it the spreadsheet, and describe what you want to achieve rather than how to build it. Have it plan first, read the plan, then let it build and test.

We had around a dozen other brilliant questions—covering everything from part-time holiday allowances and receipt scanning to donation matching—and we’re working through all of them in our Vibe Coding WhatsApp group.

Join the AI for Business WhatsApp Group – Join the conversation here.

One more example: the seminar itself

At the end, we revealed that attendees had been using vibe-coded software to run the seminar. The whole event ran on StageCast by DATAFORT.COM, a platform our team built around our own requirements. We wanted control over features that other platforms make expensive or restrictive, so we built what we needed.

StageCast covers an event from start to finish:

  • Before the event: registration and reminder emails.
  • During the event: everything needed to run a smooth session with up to 6 presenters and 100 attendees.
  • After the event: Meaningful clips cut and posted to social networks with appropriate commentary and links back to your website, and the recording emailed to attendees, to keep the conversation towards your business goals. (To find out about using StageCast for your business email: [email protected])

It’s the same lesson as the booking system. Vibe coding got us there fast. Engineering for security, integration and reliability is what made it fit to run a live event in front of an audience.

Where DATAFORT fits

Vibe coding puts real problem-solving within reach of the people who understand the problem best. We want you to succeed with it. If you’ve already started and got stuck, you’re not doing it wrong. Getting from prototype to something your business can rely on is genuinely hard.

That’s the work we do every day: understanding how a new tool has to fit with your existing systems, who should see what, and what happens when something goes wrong at nine on a Monday morning.

You know where the frustrating processes are in your business. We can help you fix them properly.