Vibe Coding vs AI-Assisted Development: What I Told the WP Remote Show

In October I sat down with Aditya Jha from WPDev for the WP Remote Show to talk through a real case of vibe coding vs AI-assisted development: a course registration system for a nonprofit, built and launched in nine weeks with AI doing much of the coding and a human reviewing every line.

The episode runs 25 minutes. You can watch it on YouTube: 5 Ways This Agency Uses AI to Build WordPress Sites Faster or embedded at the end of this post.

If you run a company and you’ve been hearing the term vibe coding, this episode is a working example of vibe coding vs AI-assisted development on a real deadline. Here’s a summary.

The deadline that would have been a no

The client runs a continuing-education program for people in retirement communities and for the surrounding community in general. Around 40 instructors teach dozens of courses each term, and registration opens online at 10:00 a.m. on one Monday every fall and spring. The first day is virtual registration. Throughout registration week, volunteers also sit at tables with laptops helping people sign up in person.

For years the site was a static WordPress build. Every course added or changed meant up to five pages to update by hand, and a single schedule change sent us back through all of them with search and replace. The multiplier increased risks of errors and made every term painful to launch.

We gathered requirements from April through June, recorded the meetings, and pulled the core requirements out of the transcripts. The client signed off on July 13. Registration day was August 24.

On a traditional development plan I would have declined and told them to run one more term by hand so we’d have time to build this. I had already experienced on another project what AI could do when a human project-manages it closely, so I promised them registration would open on August 24 at 10:00 a.m. and it would work.

The five roles we gave AI

We wrote a creative brief and a technical brief, then had the AI read both along with every meeting transcript and summary before it drafted the technical documents. From there the AI played five parts, with a person directing and reviewing at each one. That human direction is what separates AI-assisted development from vibe coding.

1. Pair programmer. The AI wrote the core features. Our developers read every line it produced and decided whether it met the requirements, whether it was secure, and whether it was something we wanted to deploy.
2. Investigator. When a bug showed up, the AI dug into it. On launch day it was our fastest way to find the cause of anything odd.
3. Deployment operator. Branch, dev, test, live. The AI ran the release steps that usually eat human hours, and we could step in at any point.
4. Advisor. Given the brief, it would tell us what it understood we wanted built before it built anything, which caught misunderstandings early.
5. Record keeper. More on that below, because it produced the most surprising number in the project.

Where the humans earned their keep

AI produces a lot of working code very fast. It has a weaker feel for how a person experiences a screen. We kept finding a button in the wrong place, a button that should have existed, or a prompt worded in a way that would confuse someone.

That gap showed up constantly on this project because the registrants are mostly (but not all) seniors who wanted a little more hand-holding through checkout. We spent real time shortening wording, cutting steps, and swapping technical phrasing for plain language. Those small edits made the site feel comfortable and non-threatening.

We also made code decisions as we went: consolidate this class, move that function somewhere else, stop doing it this way. The AI stored those preferences in memory and followed them from then on, so the corrections compounded instead of repeating.

Four quality gates: where AI-assisted development parts ways with vibe coding

Speed never overran quality or our definition of done. Every change passed through the same four gates we use on any project: human code review, automated checks (linting plus UI, security, and accessibility tests), deployment through dev, test, and live, and a real-browser check at the end.

We ran those checks continuously throughout development rather than two hours before go-live. The last two hours of any software project should be for sitting back with a drink and a snack and watching the clock for the moment we go live. That’s how it went. On registration morning the team sat around a kitchen table socializing and having fun, and at 10:00 our “ribbon cutting ceremony” was me checking the checkbox that opened registration. People kept saying they couldn’t believe how smoothly it was going.

The refund rule that slipped past us

You can run a system through thousands of tests and still miss how another person will use it. We ran registration through dozens of checks, and the one thing that slipped past all of us was a refund rule.

Everyone on the team assumed a refund meant refunding a whole membership plus courses. Once live, a registrant needed to drop a single course and keep her membership and fees intact. Our assumptions had all lined up the same wrong way.

We described her situation in plain English to the AI, said we were on a time crunch to help this individual, and asked it to find the missing case. It traced the gap, built the single-course refund path, and we had it tested and live in about 20 minutes with a person checking the output at each step. We were building the wings on development and test servers while the live site’s plane was in the air. The discipline we’d set up in the weeks before is what made that safe to do. That discipline is the whole difference in vibe coding vs AI-assisted development.

473 commits, and why that number was on purpose

The project finished in nine weeks with 473 commits, about 150 merged pull requests, and 70 releases. We shipped three days early, processed real money before launch, and the system now needs roughly an hour a week of attention, most of it deciding what to push to the next phase.

When I pulled those numbers together afterward, 473 commits stopped me. It traced back to one instruction I gave the AI at the start: log an issue for every single thing you find, resolve it, turn it into a pull request, commit it. I had wanted auditability on our first full-spectrum AI project.

The result is a granular trail of every decision from the first time a problem appeared. If something goes wrong, a person or the AI can walk the chain backward. And on the next project I can point the AI at the whole history and ask for a lessons-learned document to carry into the next build. What some folks still in the traditional coding discipline mindset may still consider “bloat” in the commit count ended up buying intelligence we can mine later.

What AI-assisted development vs vibe coding changed for the client

Every course now lives as one custom post in WordPress. Edit it once and every page that references it updates. That alone ended the copy-paste rework that had defined every term.

Payments run through Stripe, which the AI helped wire up, with the financial visibility the organization needed. Registrants told the staff the process was far easier than in past terms. The staff who used to spend registration morning putting out fires spent this one answering ordinary questions.

Vibe coding vs AI-assisted development

Aditya asked me whether clients expect these timelines now. Some do. People are starting to understand how fast software can be built, which is why vibe coding exists: skip the agency, open a tool like Lovable, and prompt until it looks about right. In a couple of years a lot more people will assume they can do that with whatever copilot they have open.

There are tiers to this. Consumer-grade AI helps with a spreadsheet. Coding-grade AI is a different animal, and if you skip the discipline it will eventually bite you. With vibe coding the prompt is the whole plan, and is only as good as the prompt itself. If the prompter doesn’t know what to ask for, disaster can occur. Usability, security, and accessibility can all break down very quickly.

But, with disciplined delivery, a human talks to the humans who will use the system, writes the brief, brings AI in as an advisor, and then reviews, lints, and tests everything before it ships. “It runs so it ships” is the gap in vibe coding vs AI-assisted development: one side prompts until it looks right, the other side has a human review, test, and approve every change before it ships.

This AI pair-coding process is now how Webidextrous builds software. Each of our developers has access to industrial grade AI to do their work. We’ve rewritten our developer and project-manager process documents around it, with AI’s help, based on what this project taught us.

For years I told clients they could have good, fast, or cheap and pick two. And that was considered the “Iron Triangle” law of software development up until now. I strongly believe AI is moving us closer to getting all three. I won’t promise it outright, because doing anything right still takes effort and still costs money, and you cannot expect the machine to get it right on its own. But we can finally take on more projects that, before AI, we never would have considered.

Sources

• 5 Ways This Agency Uses AI to Build WordPress Sites Faster, WP Remote Show on the WPDev YouTube channel, published October 17, 2026
• WP Remote, the show’s host