Know when not to vibecode
The question is not only whether AI can build something. You also need to decide whether vibecoding is an appropriate way to build it.
Capability is not the same as suitability
AI agents can help build surprisingly sophisticated systems. But a system can be technically possible to build and still be a poor candidate for vibecoding.
A useful question is:
What happens if this system is wrong?
The greater the consequences of failure, the more important it is to add expert review, stronger safeguards, formal testing, or a different development process.
Watch for high-consequence situations
Be especially cautious when:
- errors could seriously harm people or the organization,
- the system makes important decisions about people,
- sensitive or regulated data is involved,
- security requirements are significant,
- the system controls money or important transactions,
- failures would be difficult to detect,
- very high reliability is required, or
- the system will become critical organizational infrastructure.
These situations don't necessarily mean “never use AI.” They mean that vibecoding alone is probably not enough.
Bring in missing expertise
Reasonable vibecoding project:
Build an internal prototype that helps a manager explore
staffing patterns and compare scheduling options.Very different project:
Build and deploy the production payroll system that
calculates employees' pay and automatically sends payments.The second system involves financial consequences, legal requirements, security, and reliability expectations. An agent may still help, but you'd want people with the relevant engineering, security, legal, payroll, and domain expertise involved.
“Don't vibecode this alone” is often a better rule than “don't use AI.”
Your ability to verify matters
Vibecoding works best when you understand the business problem well enough to judge whether the result makes sense.
If you don't understand the domain well enough to recognize a dangerous mistake, you need someone who does.
You might be able to build a financial forecasting dashboard without being a software engineer because you understand the metrics and can verify the calculations.
Building a system that interprets complex tax rules is different if nobody on the team understands the tax rules well enough to evaluate the result.
Prototype risk and production risk are different
A useful prototype can tolerate conditions that would be unacceptable in a production system.
| Prototype | Production system |
|---|---|
| Used by a few people | Used across the organization |
| Mistakes can be noticed manually | Mistakes may happen at scale |
| Outputs are reviewed | Actions may happen automatically |
| Failure is inconvenient | Failure may be costly or harmful |
Moving from prototype to production usually requires stronger testing, monitoring, security, documentation, and ownership.
Use the four-question framework
Before deciding whether a project is appropriate for vibecoding, return to the questions from the beginning of this module:
| Question | Warning sign |
|---|---|
| What can it see? | The system requires highly sensitive or regulated information. |
| What can it do? | The agent can take actions that are costly, public, or hard to reverse. |
| What happens if it is wrong? | An error could cause serious financial, legal, safety, or reputational harm. |
| Who is responsible? | No one has the expertise or ownership needed to verify and maintain the system. |
Your work is still your responsibility
An AI agent is a tool, not a party you can transfer responsibility to. If you approve, submit, send, or deploy its work, you're responsible for deciding whether that use is appropriate.
This includes following company rules on data, security, intellectual property, and AI use. In class, it also means following your instructor's policy on AI assistance and disclosure.
A useful rule
The higher the stakes, the less you should rely on vibecoding alone.
Use agents where they increase your capability, but bring in stronger processes and additional expertise when the consequences demand it.
Exercise
For each project below, decide whether you would:
- vibecode it yourself,
- vibecode a prototype but involve experts before deployment, or
- use a more formal development process from the beginning.
| Project | Your decision |
|---|---|
| An internal dashboard for tracking marketing campaign performance | |
| A tool that recommends which job applicants should receive interviews | |
| A prototype that summarizes customer feedback for a product manager | |
| A system that automatically approves customer refunds | |
| A production payroll application |
For each choice, use the four questions: what can it see, what can it do, what happens if it's wrong, and who is responsible?