Vibe Coding Still Needs Guardrails

Situation

Over the weekend, a client I built a SaaS for tried to change the product design with AI-assisted coding.

The goal was reasonable. He wanted to move faster on the look and feel without waiting on every small design adjustment.

The failure was also reasonable, in the way failures usually are. His Claude Code agent changed the admin login password.

The result was simple: the client could not log in to the admin side of his own site.

That is not a philosophical problem. That is an operations problem.

What went wrong

The issue was not that AI-assisted coding is bad.

The issue was that the agent was working without enough operating context.

The client has been following the idea that one central “brain” should run everything. In his setup, that brain lives in an Obsidian vault. It tracks how he thinks, what he is working on, and how decisions get made.

That can be useful for planning.

It gets risky when the same brain starts working inside a code repo and assumes it understands the system.

A production SaaS repo is not just files. It has seed data, migrations, environment variables, deployment rules, authentication flows, CI checks, rollback paths, and all the little pieces that keep a business from losing a day to a preventable mistake.

In this case, the agent missed enough of that context to change something it should not have touched.

The part that worries me

Changing an admin password is bad.

It was not the first warning.

The previous week, the same pattern showed up in a PR that tried to change seed data. I denied it because seed data is not a harmless design file. It is part of how the application knows what the business expects to exist.

That was the review system doing its job.

The real problem was what happened next. The owner wanted to keep moving after hours and did not want to wait until working hours for the gatekeeper to review the change. So he tried to route around the checkpoint.

That is how a useful AI coding workflow turns into a production risk. The agent might change a migration, delete seed data, drop records, touch billing logic, or make a deployment path look clean while quietly removing the safety rails.

I have the database server set up with regular backups. That matters.

But backups are not a strategy for letting an unsupervised agent wander through production code. Backups are the last line of defense after better controls failed.

Vibe coding is not the problem

The idea behind vibe coding is useful.

A capable operator can describe an outcome, let the agent take a first pass, inspect the diff, tighten the result, and keep moving. That is real leverage.

But leverage cuts both ways.

The same agent that can update a layout can also rewrite auth logic, change seed files, bypass a test, or decide that an existing CI/CD setup is noise because it does not understand why it is there.

That is where the stories about AI coding failures stop being internet stories and start becoming client problems.

The safer pattern

Do not let the agent treat the repo like a blank canvas.

Give it a bounded job.

Tell it what files it can touch. Tell it what files are off limits. Make it work on a branch. Make it explain the diff. Run the existing tests. Run the existing build. Use the CI/CD path that already exists. Keep staging between the agent and production.

Most important: review anything that touches auth, billing, database schema, seed data, environment variables, or deployment configuration. If the person who owns that review is offline, the change waits.

Those are not design details. Those are business continuity details.

Business takeaway

AI can help a business owner move faster.

It can also move faster in the wrong direction.

The answer is not to ban the tools. The answer is to build a working system around them: permissions, branches, tests, backups, CI/CD, staging, review, and a person who knows where the dangerous parts of the repo live.

If you want that kind of setup for your own business, I handle that through Digital Spirit Technology as Custom AI Sprint work.

The goal is not to make the agent feel powerful.

The goal is to let capable people use AI without handing the keys to production to something that does not know what a bad Monday costs.