Replica turns a language description into a real, shareable GitHub prototype with Copilot. I built and designed an agent that shows new users what it can do.
Product Designer | Shipped on September 2026
My Contribution
Engineering
Product Design
Outcomes
Any hubber can now click on “What’s new?” button and stay up-to-date with Replica’s new features. This helps users take complete advantage of the tool
Shipped net-negative code in a fast-growing codebase and minimized UI surface on the platform
When agents build the UI, the design system has to speak to them too
In my second week at GitHub, a new AI prototyping tool called Replica launched. Describe a page in plain language and it builds a real, shareable prototype with real Primer components and real GitHub data. It can also pull live data, compare concepts, and hand work off to dev, and my design process soon shifted almost entirely into it. However, it was built fast and never documented. By the time I started, replica had 41 agent skills and 42 rebuilt pages, and no way to find any of them. I talked to 30 hubbers and some of them had never heard of it, some knew a little, and some used it daily without knowing its full power. Everyone had built their own process around whatever they happened to know.
I set out to answer one question: how do you tell people what a fast-moving tool can do, when the answer changes every week? Feature discoverability was already a top-3 priority for the team, and the team wanted fewer UI surfaces to maintain. I shipped a solution with product direction from Jannes Peters, Director of AI & Core Workflows Design at GitHub. One small button that opens a Copilot session, where an agent reads the repo and picks the three most relevant features right now. Quick and copilot-native.
Things to keep in mind
Reasoning
Twenty people had shipped at high velocity for 3.5 months: 4,054 commits, 41 skills, 42 routes, 48 specs. That's not a small gap in documentation, it's a structural one.
The team built all 41 things and therefore knew all 41 things, which made the gap invisible to the people who most needed to see it.
Goal
Tell a new user what a fast-moving platform can do, when the answer changes every week and nobody has time to update "what's new" page. How to maintain without maintaining?
Dynamics
It couldn't go stale. Any hand-written list will be wrong within days at that commit rate.
Nobody would maintain it. There was no docs owner, no curator role.
It couldn't add a UI surface. The team was actively trying to reduce the surfaces it maintained.
Context
Engagement was never the actual problem, people who found the platform explored it.
There was no one to be the curator, and building for one would guarantee the feature decayed by week three.
Formation
Noticing the gap as a user, not as a researcher
I'd used Replica for a few weeks and kept finding good features by accident, usually because someone mentioned one in Slack. The pain was real and current. It read as an observation worth acting on immediately. Jannes replied the next morning, agreed we had more hidden features, and the conversation that shaped the real solution. Prototype discovery already worked on its own. Capability discovery didn't, and this project closed exactly that gap
Define agent skill workflow based on trust and judgement
A system that updates itself still needs a rule for judgment, not just a way to deliver information. Automation only earns trust when users can see what it's doing and stay in control. I compared it to how an EV decides what to tell the driver during an over-the-air update. So I asked two questions: "What will the criteria be to classify unknown future features?" and "What updates should be visible vs invisible to users?" This led to the loop for the agent.
UI sparked discovery and chat added depth
The real pain point was finding out about features before finding out more about them. UI alone solves discovery but not depth. Chat alone solves depth but not discovery. So the solution to combine them both was ideal. A slight hint on the UI nudged users to discover new features from that week. The interactive agent skill brought repository context and replica context to action-oriented points that users can next try.
Progressive disclosure for detail
"Make it agentic" on its own has a hole in it. An agent can't tell you about a feature you don't know to ask about, and a blank chat box makes that worse. So, the agent lead and revealed information:
It shows the landscape first, three features at once only
It then stops and lets the user picks what to do next with action
It expands only when asked, with three how-to steps for feature
It never launches a session unprompted. It asks for consent first
