Architectural documents transforming into a glowing digital building concept in a modern studio.

From a Month of Production to a 2-Minute Draft: AI Case Study Pipeline

We developed an AI pipeline that runs entirely in VS Code (with Claude extension) on top of the bureau's project workspace and turns raw project files into a finished case study — text, illustrations, and cover.

Netherlands

Solution:

AI Content Pipeline

Services:

AI Transformation, AI, Workflow Automation

KPI:

Project Timeline: complete the project in 2 weeks.

First case draft in 2 minutes.

15 designer hours saved per case study.

Client & Context

A Netherlands-based architecture and design bureau delivers 5–10 building projects a year. We built its AI-native PM workspace in a previous project.

Case studies bring the bureau new clients. The marketer who writes them sits outside the project context. One case took weeks of research, interviews, writing, and design queue.

A Netherlands architecture studio at dusk: a detailed physical building model and rolled blueprints on a wooden table, iMacs showing 3D models, and Amsterdam canal houses through tall windows.

The bureau ships 5–10 building projects a year out of rooms like this — and until the workspace, every brief, kickoff, and set of minutes was carried by hand from one project to the next.

Goals

1
Free the marketer from manual project research.
2
Keep every published case in one structure.
3
Ship case visuals without booking a designer.

Challenges & How We Overpowered Them

Marketer writes cases from outside the project context.

Skills read the project workspace directly — briefs, estimates, scope, tickets, transcripts.

Designer booked by the week — one cover took up to a month.

Concepts, prompts, and final images now appear in the same chat in 10 minutes.

AI text tends to drift into fluff and invented numbers.

The skill enforces structure and word limits, and flags every unverified claim.

From Raw Project Files to a Published Case

One workspace as the single input

The pipeline stands on the PM workspace from our previous project. Briefs, estimates, scope, timelines, tracker tickets, and call transcripts already live in one git repo.

There is no custom software here. The whole pipeline is plain Claude with skills loaded on top.

The marketer points the skill at a project folder. Research is over.

Scattered mixed project documents on the left merging into a single glowing git-repository cube on the right, dark 3D style.

Every brief, estimate, ticket, and call transcript already lives in one repo — so "research" for a new case is no longer a stage, just pointing the skill at a folder.

A structured draft in two minutes

The writing skill reads scattered, unstructured files and produces a full draft: headline, KPIs, goals, approach, results.

Nobody writes a prompt each run. The rules live in the skill, so every draft comes out predictable — in the bureau's own structure and format.

Every KPI must carry a number. Any claim missing from the source files gets a visible flag.

Screenshot of VS Code in three panes — a blurred Imaga project workspace file tree on the left, a generated case study draft in the center editor, and an agent chat with prompts on the right.

No research folder to assemble and no second app: the project repo sits open on the left, the draft takes shape in the middle, and the marketer just talks to the agent on the right — the whole case is written without leaving one window.

Validation replaces writing

The marketer checks facts against reality and confirms the structure.

Editing and unifying the format used to be a separate stage. Now the structure ships built into every draft.

Claim cards moving from a document through a gate labeled "validation"; two cards are blocked and pinned with flags.

The human job moved from writing to checking: every claim passes through a gate, and anything the source files can't back up is flagged before it can reach a published case.

Image concepts from the finished case

The second skill reads the approved case and finds its most visual moments.

It walks the marketer through gated choices: style options with moodboard previews, an image map, then five concepts per image with layered AI prompts.

Diagram of three connected modules labeled writer, image-concept, and cover along a central spine, each emitting small output cards.

Three skills split the job — one writes the draft, one builds the image concepts, one designs the cover — and none of them hands off to a separate tool or person.

Images render in the same space

Generation happens right where the concepts appear. The marketer reads a concept, approves it, and watches the image render in the same chat — if works in the Claude desktop, and gets links — if works in the VS Code.

This works through Magnific, connected to Claude over MCP. Magnific is an AI image-generation platform. MCP is an open standard that lets Claude use outside services straight from the chat.

The handoff to a separate tool or person is gone.

Two nodes labeled Claude and Magnific joined by a glowing arc bridge labeled MCP, cards traveling both directions inside one outlined space.

Concepts go out and finished images come back across the same MCP link — which is why the picture appears in the chat the marketer is already working in, not a second tool.