What We Would Have Built Next

The end of thirteen weeks

Thirteen weeks goes faster than it sounds. Both tools are built, tested, and in the hands of real users. And now that the placement is over, the most honest thing we can do is look back at what we built, what we didn’t get to, and what we would have kept going on if the clock hadn’t run out.

This is that conversation.

What we would have kept building on the web side

Phase one proved the pipeline. A structured brief goes in, a complete website comes out — with real brand identity, real SEO foundations, real quality standards built in by default. That works. The question that didn’t get answered is what it would have become with more time.

Image generation was the most obvious remaining gap. Claude’s image output isn’t good enough for hero sections and visual-heavy pages — Gemini handles that significantly better, and integrating it directly into the pipeline was the clear next move. It just didn’t make it into the timeline.

Versatility was the other thing. The pipeline generates full sites. What it couldn’t yet do was generate a single page, or a single section, without rebuilding everything around it. A client who needs a new landing page shouldn’t have to run a full site build to get it. That’s a real constraint that more time would have addressed.

The generation experience itself was also worth rethinking. The output lives in a code editor — functional, but not what a non-technical client wants to look at. Moving generation onto the webpage directly, so what you see is what you get, was always the right direction. It’s expensive to build, which is why it stayed on the list rather than making it into the build.

Hero video is currently handled through a third-party step. Building that in natively — brand-specific prompts, automated generation, directly integrated into the build — would have removed a manual handoff that shouldn’t need to exist. Drag-and-rearrange editing built into the page, and automatic form connections and backend setup, were the two features that would have closed the biggest gap between what Web AI produces and something a client can fully own without developer involvement.

What we would have kept building on the image side

Phase one moved fast. That was the right call for the timeline — but moving fast leaves marks, and with more time the first thing that would have happened is paying those down.

Some pieces of the codebase ended up doing more than they should. Logic got duplicated in places that then drifted apart. Shortcuts that were load-bearing in week three were still load-bearing in week twelve. None of it stopped the tool from working — but it’s the kind of thing that makes the next person’s job harder than it needs to be, and that quietly slows down every change that comes after. A proper refactor was on the list. It didn’t make it in.

Beyond that — more editable shapes and vectors, deeper AI understanding of brand assets, and a handful of features the team asked for during testing that were genuinely interesting but too complex to fit in alongside everything else. The model usage logging that went in during week eleven would have informed a lot of this — looking at where costs were accumulating, where the system was defaulting to expensive models when cheaper ones would do the same job, and using that as a signal for where the pipeline assumptions didn’t match how people were actually using it.

The tool got to a good place. With more time it would have gotten to a better one.

What both projects are leaving behind

Both tools are leaving behind something more than a codebase. They’re leaving behind a way of thinking about how AI fits into agency work — not as a replacement for creative direction, but as infrastructure that does the execution so that direction can be the focus.

The brand intelligence framework that both tools share is probably the most transferable thing from this co-op. The idea that brand context should live in the system, load automatically, and constrain the output before a user has typed a single word. That idea scales. It applies to tools that don’t exist yet. It’s worth building on.

Whoever picks this up next starts from a much stronger position than we did. The pipelines are real. The users are real. The feedback is real. That’s not nothing — that’s most of the hard work.