Building a tool to resolve implementation questions before release
An internal AI tool that helps teams resolve implementation questions, move faster, and build consistent interfaces and code. It brings component previews, code, and guidance together while helping designers learn how the design system works.
At a glance
Ask Atlas gives designers, engineers, product managers, and non-technical teammates a way to ask what already exists, what can change, and when they need to build something new. Engineers can ask what a component supports and check the recommendation against its code before making an implementation decision.
I saw an opportunity to help the team spot and fix inconsistencies across products, speed up quality assurance (QA), and give designers practice reading component code. The goal was to spend less time tracking down answers and build a shared understanding of how to use and maintain the design system.
Component decisions became release problems
Tricura shared a design system across three products. A question as simple as “Can this button have a different corner radius?” raised a bigger decision. Could the team change it on one screen, or would the change affect a component that other products relied on? Each round of clarification took time away from design and engineering, and unresolved questions made it harder to implement changes consistently.
- Design and engineering
- Questions about component settings came up regularly in the design channel and during reviews. Engineers spent time tracing inconsistencies and explaining what the components supported.
- Stakeholders
- Even small component decisions could take several rounds of discussion to clarify what the shared system supported.
- Product
- Local overrides could make the same component look different across products that shared a design system.
- QA and release
- The QA team had to ask how components and interactions were supposed to behave before they could verify them. Questions left open at handoff could resurface after release, when the team had to trace and correct the implementation.
A workshop turned recurring questions into a clear scope
I ran a workshop with designers and engineers to review past cases and look for patterns in the questions that kept coming up. Designers described how they had to look in several places to inspect a component. Engineers brought questions they encountered during handoff and implementation.
We shaped the tool around a single place to inspect components and answer three questions. Do we already have it? Can I change it? Should I build it?
One answer shortens the path from question to handoff
Ask whether an overdue label can be red and the tool shows the Badge component and the setting to use. Ask whether a button’s radius can change and it explains why the pill shape is shared and which settings can change. Ask whether to build a medication schedule and it suggests combining a table, status badges, checkboxes, and row actions using our existing table patterns.
Depending on the question, the tool provides a recommendation, information about a component, or implementation guidance. Designers can try the component and copy the answer and code into the handoff or send them directly to engineers. Engineers can ask their own questions as they build, giving both teams the same reference for component settings and code.
Checking AI answers against the code
I initially proposed using the Anthropic API and briefed leadership on the opportunity. API access wasn’t readily available, so I built the proof of concept as a Claude Artifact. That gave me a way to test the approach and bring the findings back to leadership.
I connected the tool to a catalog built from the Atlas codebase and the three product repositories. I also brought in product guidelines and reused selected skills from a separate design system agent I built. Figma served as the reference for how design variants map to code.
Learning to read code
The tool was an opportunity to build the team’s skills through everyday work. Some teammates wanted to learn to code, and I wanted to help them read component code, understand the design system, and check AI suggestions. When a designer changes a button’s size, the preview and code update together. They can see what a prop, or component setting, does and use that knowledge in a conversation with engineering.
Bringing answers to Slack
The workshop surfaced an opportunity to bring Ask Atlas into our main Slack channel. The team already has a process for asking and discussing these questions there. A Slack integration could answer routine questions in the thread and link to Ask Atlas when someone needs to inspect a component or try its settings.