There's a lot of talk right now about "AI-powered development." Most of it stays vague.
A demo here. A chatbot widget there. A blog post about "the future of coding."
We wanted to write something more concrete instead: what it actually looks like to build a production system with Claude, end to end, for a real paying client.
At InterCode, we're an AI development studio. Our team is made up of CCAR-F Claude Certified Architects by Anthropic.
That means Claude isn't a side experiment for us. It's part of how we architect systems from day one.
In practice, that usually means Claude shows up at more than one layer of a project. Not just as a chatbot bolted onto a finished product afterward.
It's part of how the system gets built. Part of how it makes decisions once it's live. Sometimes even part of how it looks.
We've done this enough times now to have a pretty clear picture of where each layer actually earns its place. Here's a breakdown of one project where all three show up.
Where Claude fits into the stack
We don't bolt on "an AI feature" as an afterthought. On this project, Claude runs through three distinct layers:
1. Claude Code — for building the thing itself Our engineering team used Claude Code throughout development: designing the microservice architecture, working through the data pipeline design, and writing backend and frontend code. This is the layer most engineers reading this will recognize - an AI pair-programmer embedded in the actual dev workflow, not just an autocomplete suggestion.
2. Claude API — for the product's own intelligence This is where it gets more interesting. The Claude API powers several in-product features that the client's marketing team now relies on daily:
Turning property specials into optimized ad copy automatically
Classifying and excluding negative keywords, using property-level context instead of generic keyword lists
Ingesting emailed reports and alerts to surface anomalies and recommended actions - without a human reading every report line by line
Powering an internal natural-language chatbot so the marketing team can just ask the data questions instead of building a new dashboard filter every time
**3. Claude for the UI/UX design work **Even the design side of the platform leaned on Claude — from early UI exploration to refining the actual dashboard experience users interact with daily.
The point isn't "we used AI everywhere for the sake of it." Each layer solves a specific, boring, expensive problem: writing ad copy by hand doesn't scale, reviewing keyword lists by hand doesn't scale, reading every alert email by hand doesn't scale. Claude removes the manual step at each of those points, with engineers still supervising the output.
This particular project is a BI and marketing-automation platform we built for Street Digital, a digital marketing agency serving luxury apartment communities — replacing a no-code setup that worked fine early on but couldn't keep up once their client base grew.
Why this matters beyond one case study
Most companies don't need "an AI feature" bolted onto an existing product.
What they actually need is someone to look at the real bottlenecks. The manual ad copy. The manual keyword review. The manual report reading.
Then figure out which of those a properly supervised AI agent can take off someone's plate entirely — and which ones still need a human in the loop.
That part is genuinely hard to get right.
It's where having architects who understand both the engineering side and Claude's actual capabilities makes the difference. Not treating Claude as a black-box API call. Treating it as part of the architecture.
Top comments (0)