Understanding the challenge
MTF was exploring AI-assisted design tools for MTF Connect, a compliance-sensitive lending platform. The risk was specific: a generic AI confidently generating a "plausible" screen is genuinely dangerous in a regulated workflow – it won't know a decision needs an audit trail, or that a status is system-generated rather than a human call. Making the design system AI-ready needed to mean something concrete, not a buzzword.
My role
I designed and built this AI-context package myself – defining the guardrails, the design-thinking framework, and the mandatory review gate – then packaged the whole thing as a real, invocable Claude Skill.
Before any design rule, the package states what MTF Connect actually is: not a consumer banking app, an operational lending platform. From there, I built a Jobs-to-be-Done framework of 10 real jobs, six working design principles, and explicit "Always / Never" guardrails – backed by a governed component library of 49 components and 198 design tokens sourced directly from the real NUI Figma file, so nothing gets invented on the fly.
I wrote down an explicit priority order rather than leaving "what matters most" implicit: accuracy over speed, compliance over polish, transparency over automation that hides its reasoning. A mandatory 15-point review gate, including a dedicated hallucination check, slows the process down on purpose – in a regulated lending platform, a screen that's fast but quietly wrong costs far more than one that's slower and correct.
Packaged as a real Claude Skill that loads context automatically:
– A 15-point Design Review Gate ending on one test: could an Originator immediately understand what's happening and why?
– A coaching layer offering grounded suggestions throughout the process, not just one big review at the end
– A Design Outcomes file closing every project with a named sign-off across Design, Engineering, Product, and Compliance.
A validated pilot: an application header component, built entirely from NUI context, proven across all five application types – it had to survive real lending logic, not a demo scenario. That validation moved the system to its next stage: it's now being used to prototype whole new MTF Connect features, starting with Payment Waiver and Hardship.
