Claude AI: Stop and Look Before You Vibe Code
The video above is a working session, not a lecture. It is part of my Claude for software testers series, and in it I sit down and prompt Claude to apply a testing-first approach before any vibe coding happens. Below is the written version, expanded into a fuller guide.
Watch the full video on YouTube: Claude for Software Testers, Vibe Coding the Right Way.
Vibe coding is what happens when you describe what you want and let the AI build it, and it is spreading fast among people who do not know much about development and even less about testing. In the video I open with a real example: someone had vibe coded a project, it was clearly broken, and Claude kept insisting it worked fine. We had to step back and actually verify it. That gap, between what the AI says it built and what it actually built, is the whole reason testers matter here. This guide expands the session into a method you can repeat: how to make testing the first move instead of the cleanup crew.
The people shipping vibe-coded apps are not going to test them. You are. That is the opportunity.
Why Testers Have the Edge in a Vibe-Coding World
Testers know what is going to work and what is not, which is exactly the judgment vibe coding is missing.
The challenge with vibe coding is not that the AI cannot produce code. It is that the people prompting it often cannot tell good output from broken output. Claude will tell you the build is fine while the build is on fire, and someone without a testing background will believe it. As a tester you have the one skill that does not come in the box: you know how to confirm whether something actually works. That makes you valuable in two directions at once. You can build with these tools yourself, and you can educate the people around you who are creating AI applications without any testing instinct at all.
Apply Testing First, Before You Vibe Code Anything
Before laying in requirements or architecture, get the AI to commit to a testing approach first, then build.
This is the heart of the session, and many of you will recognize it as TDD by another name. On the applications I have built through vibe coding, I make Claude apply a software testing approach before it writes the thing.
In the video I do exactly that out loud: I tell Claude that before any coding, we want everything thoroughly tested, with unit, integration, and end-to-end activities planned, a test strategy and test plan in place, Playwright lined up for API and end-to-end coverage, and a master test prompt that bakes in the best practices. Test first, then develop. On apps I have built this way, I found that when you set that expectation before the build, the AI organizes the whole project around testability instead of bolting it on at the end.
Make Claude Ask Questions, One at a Time
The most useful instruction in the whole session is telling Claude to ask clarifying questions before it does anything, and to ask them one at a time.
When I tried letting it run without that instruction, Claude rushed off and built on its own assumptions, like most large language models do. So I tell it up front to review the request and ask me anything that would help before it starts.
In the video it comes back asking about audience, the depth of AI testing, who runs the master test prompt, the technology stack, and the quality gates that will block a merge. I also tell it to ask those questions one at a time rather than dumping six at once, because answering them in sequence is far easier than trying to hold all of them in your head. That back-and-forth is not friction. It is the AI loading the context it needs to give you a usable result instead of a generic one.
Let the AI Audit Its Own Plan
When I asked Claude to check the plan for gaps before building, it found twenty-three of them.
Even with a video plan already written, I had a feeling things were missing, so I told Claude to review the whole plan and tell me what we would have to fill in later. It came back with twenty-three gaps, which is significant. Test data management, exploratory testing, performance, security, accessibility, CI/CD, debugging and reporting, and testing AI systems were all missing.
It also flagged the things that bite you in practice: end-to-end tests built around tools like Selenium will be flaky, the testing pyramid needs the right balance with more unit than integration than end-to-end, mock services for when you cannot use real data, a clear definition of done, and naming and documentation standards. Asking the AI to critique its own plan before it executes is one of the cheapest, highest-value moves you can make.
Standards, Stack, and Skills Over Agents
Consistency is the point, the stack stays Claude-only to control cost, and skills often beat agents.
Standards are something I reinforce with every team I lead. Every defect should look the same: consistent titles, descriptions, steps, screenshots, and logs. I want the AI held to that same standard, so the output is uniform every time.
On cost, I pushed Claude to keep everything Claude-only with no OpenAI and no recurring spend, and it landed on a free, open-source stack around Playwright. The app we are building across the series is Risk Radar. And when I floated turning the testing components into AI agents, Claude pushed back and recommended skills instead, because skills are teachable on camera, free, and reusable, while agents should be used sparingly. That instinct, reach for a repeatable skill before a standing agent, carries through the rest of the series.
Final Thought
Vibe coding is not going away, and most of the people doing it cannot tell whether what they shipped actually works. That is your job and your advantage. Make testing the first move, force the AI to ask questions before it builds, have it audit its own plan, hold it to real standards, and prefer a reusable skill over a one-off agent. Do that and you are not just keeping up with vibe coding. You are the person who makes it safe to ship.
The full session plays out end to end in the video. Watch it above, and tell me in the comments: what is the worst thing you have caught in a vibe-coded app that the AI swore was working?
