Can vibe coding work for a project that has no close existing example?
Asked by
MellowOrbit42
I've seen demonstrations where AI is used to quickly build familiar applications, such as a movie database with a fairly standard web interface and database structure. That made me wonder how well this approach works for less conventional projects. My current project is an embedded, distributed sensor and alarm system. It may not be groundbreaking, but I haven't found a public project that closely matches it. Is vibe coding still practical in that situation, or does it mainly work when the AI can imitate established patterns and examples?
Answers
It tends to perform better when the project follows a mature framework or familiar architecture. I’ve seen it do well with conventional admin-panel customization, but struggle with less standard designs where the important architecture is harder to explain. For a novel system, use it feature by feature: write small specifications, define interfaces and constraints yourself, and review every result rather than asking for the entire system at once.
Reply to this answer
Most new software is still made from familiar building blocks: logging, metrics, interfaces, data conversion, storage, networking, and validation. AI can generate a lot of that even when the overall product is unusual. The difficult part is usually the small portion that contains the actual new idea, and that is where a human needs to define the design, constraints, and correctness criteria.
Reply to this answer
Vibe coding is most viable when quality requirements are modest: prototypes, throwaway tools, non-critical applications, or projects where mistakes are easy to spot and fix. For something safety-sensitive, distributed, or intended for long-term maintenance, you need strong tests, clear specifications, and someone who understands the generated code. AI can still help, but it shouldn’t be trusted as an unsupervised designer.
Reply to this answer
You don’t necessarily need a matching public example. What matters is whether you can describe what the system should do and provide a way to verify it. Even for an unusual embedded project, AI can help turn a clear design into code and accelerate routine work. The less familiar or more safety-critical the behavior is, though, the less appropriate pure ‘vibes’ become.
Reply to this answer
An LLM can combine known techniques into something that is new for your particular use case, but that is different from inventing and validating a genuinely new algorithm. If your project requires novel control logic, complex mathematics, or guarantees about reliability, don’t expect a vague prompt to produce a finished solution. Explain the intended behavior precisely, ask for small implementations, test them against known cases, and treat the AI as an implementation assistant rather than the system architect.
Reply to this answer
Add your answer
Answers are reviewed before publication.
MellowOrbit42: That makes sense. The sensor and alarm logic is the part I’d design myself, while AI could handle more routine pieces such as configuration tools, logging, and data formatting.