Case study 03 · In progress
AI-Assisted Frontend Implementation (Claude Code / PR)
Role — Design, frontend implementation
Product — Enterprise Admin platform
Status — Waiting on backend
Context
I began moving beyond the standard design-to-handoff model by using Claude Code directly. Starting with low-risk layout and localization updates, I gradually took on work that removed an implementation handoff step and built trust with engineering — eventually contributing complete feature pages.
The problem
Learning to work in a codebase without treating code as a black box.
I do not read code fluently, and technical debt made even small changes more complex than they appeared. In an early change, I adjusted a button used in one place without realizing its style came from a shared class — and the change affected buttons across the product.
The decision
Start small, verify every change, and let review remain the quality gate.
After the shared-button incident, I used Claude Code to understand the relationship between shared classes and instance-specific styles, then rebuilt the fix without affecting other buttons. I treated early mistakes as low-cost learning opportunities rather than aiming for perfect results on the first attempt.
Every PR is reviewed as a diff in GitHub. When I cannot follow a change, I ask Claude Code to explain the behavior rather than merging blindly. Engineer and PM review remain mandatory before merge — a guardrail that made deeper experimentation possible.
Execution
Case: Microsoft Teams integration
I implemented Microsoft Teams as a new integration option on an existing Slack and email integrations page. I reused the shared integration logic where possible, while building Microsoft-specific steps — organization selection after authentication and a six-digit bot code — from scratch.
20 files · +3,109 / −34
Built on existing implementation
Before entering the full Admin repository, I built a lightweight HTML prototype and walked stakeholders through it live. This exposed multi-select needs before code-level work began. The near-production PR then surfaced loading states and multi-organization linking — requirements that static Figma screens had not revealed.
Outcome so far
The implementation is ready for the real backend — and the learning is already concrete.
The feature has not shipped yet: the related Slack and email integration work is still in progress, and Teams is waiting on backend implementation. I am not claiming a ticket-volume or engineering-time outcome before the feature is live.
What the process already proved
An interactive prototype surfaced loading and multi-organization requirements before they became rework. The working PR gave stakeholders a more useful object for alignment than static handoff files. Once the integration ships, I plan to add ticket and engineering-time data.
Beyond one feature
Teams is one example. In practice, I have kept shipping PRs continuously — from layout fixes, to localization, to full new features. The real point is that this is not a one-off experiment; it has become part of how I actually work.