I love vibe coding.

It has changed how I work with AI—not only because it helps me make things, but because making something together often reveals possibilities I would not have found by thinking through the problem alone.

I can begin with an incomplete idea, describe what I am imagining, react to what appears, and follow the most interesting thread. I may discover a better way to structure the work. I may see a connection between pieces of information that I had previously considered separately. Sometimes the prototype gives me an answer. More often, it gives me a better question.

That is not a failure to plan. It is the point.

But vibe coding is not the same thing as automating a process. The two can use similar tools and may even begin in the same conversation, but they operate under very different conditions—and make very different promises.

Discovery mode

Vibe coding is a form of exploration. I do not necessarily need to know exactly where I am going before I begin.

In my research work, for example, I saw that an existing longitudinal evidence base could potentially answer broader questions than the recurring reports built around it. I did not start with a complete product specification. I started by exploring relationships among different dimensions of the data and using AI to help me make those relationships tangible.

As I built and reacted to early analytical and dashboard prototypes, the possible product became clearer. So did the questions it might help someone answer. The act of building was part of the thinking: each version made it easier to see what was useful, what was missing, and where the idea could go next.

I have had a similar experience building Oceane's Echo, my art and jewelry business built around original work and locally foraged or reclaimed materials. AI has helped me explore how original artwork might translate across products, experiment with visual directions, develop the store experience, improve product storytelling, and identify decisions I had not realized I needed to make.

Those explorations did not always begin with a specification. Sometimes I needed to see an idea before I could evaluate it. One product concept led to another. A visual experiment raised a merchandising question. Working on a listing exposed something that needed to be resolved elsewhere in the customer experience.

This is what I value about vibe coding: it supports discovery before the destination is fully known.

Reliability mode

Automation is different.

When a process is expected to take an input and produce the right output reliably and repeatedly, it creates an operational promise. A successful experiment—or even a successful prototype—is not enough.

In recurring research production, the opportunity is much larger than automating a file-level quality check. The real challenge is connecting the entire production system: potentially beginning with an automated export from a source platform and continuing through data transformation, reconciliation, analysis, narrative, formatting, quality assurance, and a coordinated package of client-ready deliverables.

The possibility of removing manual steps does not reduce the need for expertise. It increases it.

To automate that chain responsibly, I need to understand both the process and the data. I need to know:

  • How the source data is structured and how that structure can change between cycles
  • Which calculations, classifications, and business rules are applied along the way
  • How different metrics relate—and which apparent inconsistencies are legitimate
  • Which fields and decisions feed each downstream deliverable
  • Where an error introduced early could propagate through multiple outputs
  • Which exceptions can be handled systematically and which require investigation
  • Where interpretation and research judgment must remain human

I also need to understand why each deliverable exists.

A tracker, a written report, and an executive presentation may draw from the same underlying evidence, but they do not perform the same function. They organize information differently, answer different levels of questions, and place different demands on the audience. Producing a complete client-ready package is not a matter of transferring numbers from one file to another. It requires preserving meaning, applying the correct analytical logic, maintaining consistency across outputs, and translating evidence for the decisions each format is meant to support.

The workflow also needs controls. What initiates it? How does it know the correct files have arrived? What happens when a field is missing, a label changes, or an expected record appears twice? How is the output validated? Where does the system stop and ask for a person? What documentation would another researcher need to understand, review, or maintain it?

Without those answers, the code may run. The files may open. The numbers may even look plausible. But the system has not reproduced the work. It has reproduced some of its visible artifacts.

From creative possibility to a working product

The same distinction appears at a smaller, more personal scale with Oceane's Echo.

Exploring how my artwork might live on a new product can be open-ended. I can try compositions, rethink colors, compare mockups, and change direction when something unexpected works better than the original idea. That is discovery mode.

Preparing a collection for sale requires reliability mode.

Each product needs the correct production file, placement, garment and color choices, pricing, description, care instructions, size information, listing imagery, fulfillment connection, and publication status. A compelling mockup is not enough. The pieces have to remain consistent, the production details have to be accurate, and nothing should become visible to a customer before it has passed final review.

AI can assist throughout that sequence. Parts of it may eventually become automated. But the workflow only becomes dependable when the inputs, rules, exceptions, approvals, and expected outputs are understood.

Vibe coding can help me discover what the collection could be. Operational knowledge is what allows me to deliver it responsibly.

Both modes matter

I do not want to lose the freedom of vibe coding by forcing every early idea into a rigid process. Its openness is what makes it useful. It expands what I can imagine and gives me new ways to interrogate a problem.

I also do not want to mistake that creative momentum for operational readiness.

Before an experiment becomes a repeatable workflow, there is a transition: from discovering what might work to understanding why it works, when it does not, and what other people will depend on it to do. That is where process mapping, data knowledge, validation, documentation, governance, and human judgment enter the picture.

Sometimes understanding the process shows that it should be automated. Sometimes it reveals that only part of it should be automated. Sometimes it exposes a process that should be redesigned—or abandoned—before any technology is added.

Vibe coding changes what I can imagine. Process automation changes what I can reliably deliver.

I value both. I just do not confuse them.

And when automation extends all the way from a source-system trigger to a final decision-ready package, the need for process and data knowledge does not shrink. It grows.