The Challenge
Creating the Content Matrix was one of the most repetitive and time-consuming steps in the content workflow. It acts as the single source of truth between design and data entry: every page's copy, microcopy, navigation labels, links, and media specs must be structured there before anything reaches the CMS. In practice, this meant manually extracting everything from Figma into a spreadsheet, page by page. The process was slow and error-prone: copy-paste mistakes, inconsistent naming, and missing fields were common, and it pulled designers away from higher-value work.
The Solution
A redesigned website content workflow addressed this by introducing an AI-assisted pipeline that generates the Content Matrix directly from Figma via an MCP server. I evaluated multiple AI models and workflows across OpenAI and Claude, and selected Claude Opus for its reliability and extraction quality. The pipeline works in two steps: first, a prompt extracts copy, navigation labels, links, media specs, and estimated character limits. Then, a second step handles multilingual translation and structures everything into a spreadsheet. Designers review and validate the output before it's shared with clients for content editing and translation, or passed to whoever handles CMS data entry.
The Results
50%
Faster Matrix Builds
16h
Saved per Medium Website
2 Clicks
To Extract and Translate
Full
Team Adoption
The Process
Understanding the workflow before changing it
Designing for reliability, not just automation
From manual copy-paste to a team-ready workflow
The strongest solution was not the most disruptive one, but the one the team could trust, run, and improve.
Wrapping up
This project sat between workflow design, research, and AI experimentation. It taught me that improving a live process means designing within real constraints: budget, learning curve, available tools, and the habits of the team who will use it. The best solution was not the most disruptive one, but the one that made the work faster, more reliable, and easier to adopt.
The strongest takeaway was the value of first-hand feedback. Testing different models and prompt structures helped define the direction, but the workflow became usable through review and iteration with designers and developers.
Paths worth exploring
Although the project is closed, it surfaced approaches worth carrying into future work: a hybrid automation with a deterministic backbone and LLM steps only where language or reasoning is needed, or an n8n flow that automates more of the process while keeping designer review as the final quality checkpoint.
Credits
