How today’s vibe coders fight model drift

Like clouds, balloons, and that story you told at the company holiday party, AI models drift.

Employees across all industries are relying on vibe coding via plain-language prompts. While a global 2025 Stack Overflow study found that almost 12% of developers admit to “vibe coding” in their daily work, non-tech employees are utilizing chatbots to code, as well.

There’s a significant downside to vibe-coding’s speed, however: a careless prompt could allow an AI-powered app to act beyond its original intentions, producing wrong outputs or breaking at a crucial moment.

We spoke with creators of vibe-coded CRMs and coffeebots about how they make sure their AI ideas don’t sail off into the distance.

What is model drift? Stanford’s Human-Centered Artificial Intelligence team defines “model drift” as “when a machine learning model’s performance degrades over time because the real-world data it encounters has changed from the data it was originally trained on.”

For example, as IBM notes, new data might be inserted in imperial measurements, baffling the underlying model trained on metrics.

Promptable coding tools like Cursor and Lovable offer local project guidelines, customized instructions written by the vibe coder and not pulled from a data source on the junk-filled internet. System prompts, which can be attached as files or even placed at the top of every new query, offer starting-point rules (like how data must be structured) that can anchor a meandering model.

The customer (platform) isn’t always right. Christopher Lee, co-founder of Local Blueprint, a marketing and tech services company for the home services industry, helped a roofing company vibe-code a customer relationship management (CRM) tool.

But a vibe-coded CRM isn’t built in a day: Lee had to test versions, add updates, and then test that those updates didn’t upset the model’s reasoning. Lee works with many industries, each with their own data structure; he knows that an LLM trained on plumbing info may produce unexpectedly different results when suddenly connected without guidance to roofing data sources.

To fight that drift, he reviews three major areas following any changes:

• The user interface. Did a button disappear? Is text scrollable?
• The database. (Lee will “skim” the DB to note unexpected data structures.)
• Business process. Is the desired objective achieved?

And the answer to drift, he told us, is specificity. An open-ended prompt like “create a table where I can store different types of shingles” won’t cut it. Lee’s prompts incorporate the specific rows and columns required, along with potential variables (for example, pricing, manufacturers, or color). He’d even create a sample table for the LLM’s review in the desired format.

“Vibe coding does not mean that you can just type in ‘build me a CRM’ and then just be done with it. You have to be very particular about exactly what you’re trying to achieve and what your requirements are,” Lee said, adding that he also keeps his prompts “very focused on one business area with one business outcome.”

During the development phase, Lee said, he also uses multiple coding tools, including Codex and Claude Code, to compare outputs; after that, he has regular security and code checks.

Following any change to the code or data, he asks users to “break” the app. Employees may find, for example, that a button doesn’t work on mobile, or a new data type leads to an error calculation. Following such discoveries, Lee re-prompts the model, often by writing a new field validation rule, or providing a copy-and-pasted reference to the bug in question (or screenshots and an explanation of what needs fixing).

A typical software development life cycle features many important steps before “deployment,” including tasks like defining requirements and testing. Lee believes that a vibe-coded app still requires the same practices, and he recommends that prompts never lead to immediate production code. To that end, he has a QA test environment, and often asks the AI to provide a detailed design before doing any official building.

Good vibes? In June, IBM described how an increase in vibe coding causes a spike in security issues, including data exposures and hardcoded passwords.

Yet plenty of IT pros are vibing out despite the risks. Mike Toole, director of security and IT at cybersecurity platform Blumira, has used prompts to build his company’s internal Coffeebot—a coworker meetup tool that has been online for about a year now.

Toole’s maintenance relies on simple measurements. He sees the monitoring for HTTP response status codes as an easy win—200 means a successful connection and 400 means not so much. Analysis can be done via web server logs and firewalls.

And there’s always one metric that clearly shows if an app is working properly, or if it’s drifted from the user’s workflow entirely: engagement.

“If people stop using it, we’ll probably shut it off,” Toole said.