Making Design & UX research AI first
Power Platform Design Research | From designing screens to owning the workflow behind them
Power Platform Design Research | September 2026
We had an AI accelerated prototype generation, but it did not solve where work lived, how it followed the design system, or how another person could review it with confidence.
I designed the operating model behind the screen: repository structure, focused agent skills, intake, previews, and visual controls that returned changes to source. My role spanned product design, systems thinking, and UX engineering, with ownership of framing, architecture, constraints, and quality.
Year
09.26
Impact
Timeline
Jan - September 2026
AI could generate a screen. I designed the system that made the work repeatable, reviewable, and ready for the next designer.
I connected repository structure, design-system guidance, intake, automation, and Studio Controls into one AI-first workflow for Power Platform Design Research.
The end-to-end workflow: frame the work, route the request, build, tune, and review.
The shared repository gives prototypes a clear home and keeps guidance, tooling, and delivery paths together.
Opportunity
The screen was only one step
Generative AI could create a first pass, but designers still had to decide where it belonged, which foundation it used, how it ran, and which version was ready to share. Without a common path, local edits disappeared, review links pointed to the wrong branch, and design-system guidance stayed trapped in documentation.
I designed the path before connecting the tools
The sequence became clear: frame the work, route the request, build an isolated prototype, tune it, and review it through a draft pull request. Designers owned intent and quality. The agent handled repeatable mechanics, while the repository carried conventions, procedures, metadata, and review expectations.
Backlog Watch compares product direction, engineering activity, and design coverage before producing a shared advisory.
Design Intent
One foundation, without one giant app
Each prototype had a clear home without forcing every experiment into one dependency graph. React projects kept their own package manifest and lockfile, while simpler prototypes stayed lightweight. Common themes and helpers reduced duplication without expanding the blast radius of a change.
Guardrails at the point of decision
I translated design-system knowledge into focused agent skills that found the nearest token, compared current and requested values, and made exceptions explicit. Review guidance flagged raw colors, deprecated tokens, missing labels, and inaccessible patterns while leaving the final judgment with the designer.
Small, relevant context
A 41-line entry point routed tasks to a 925-line reference only when deeper guidance was needed. Session-start context supplied the active branch, prototype catalog, local server, and pull request state, giving the agent enough information to act without loading the entire repository.
GitHub Projects connects design intent to ownership, priority, status, surface, and delivery context.
Studio Controls keeps the selected element visible while exposing the values a designer needs to inspect and tune.
What I Owned
Planning connected intent to implementation
I created a GitHub Projects board that connected design work to ownership, priority, surface, journey, delivery date, and handoff state. A Design Intake form captured the problem, expected outcome, and relevant context, while manual triage remained visible where automation stopped.
Automation supported alignment
Eight native workflows handled item, issue, pull request, and archive events. Backlog Watch and Scrum Recap ran on four primary and catch-up schedules, comparing product direction, engineering activity, design coverage, and team discussion. Duplicate checks limited repeated reports, while people retained responsibility for priority and ownership.
Direct visual iteration
Studio Controls kept the selected element visible while exposing relevant spacing, color, typography, corners, and prototype-specific values. Browser exploration remained temporary, so handoff captured the target, before-and-after values, and source context for a durable code change. Advanced direct control was demonstrated in a working branch, not presented as a completed rollout.
A clear preview and review path
One local command opened the gallery and every prototype. Draft pull requests produced reviewable copies, while production publishing followed a separate route. When GitHub Actions was unavailable at the organization level, delivery moved to Azure DevOps without changing the collaboration model.
Impact
Verified results included 75% less setup time, 60% fewer correction rounds, 60% fewer model tokens per task, and a 30 percentage-point increase in first-session completion.
The supporting system footprint included 187 indexed prototypes, 181 native Project workflows, 4 schedules across 2 report families, and 1 predictable local development entry point.
Reflection
What changed in my practice
This project moved my responsibility beyond the screen. Repository structure, persistence, agent context, and team operations became part of the experience because they determined whether an idea could survive review and reuse.
Principles I carry forward
Put standards where decisions happen. Designers should not need to translate a reference library into every prompt.
Separate exploration from persistence. Direct manipulation helps people judge visual choices, while explicit handoff protects the source of truth.
Keep ownership and evidence connected. Boards, issues, code, previews, and reviews answer different questions, but they should tell the same story.
Adoption and operating signals
Of 31 roster-matched contributors, 27 returned in another week, an 87% repeat-participation rate.
Weekly activity averaged 8.5 active contributors in the latest four weeks, compared with a peak of 14.5. Median review turnaround was 15.2 minutes in the latest four weeks, up from 2.5 minutes in the first four weeks.
