The problem
COE UI spans Angular, Ionic, Flutter/FlutterFlow, native iOS and native Android, across three sub-teams. Everyone had access to AI coding assistants, but each engineer was working out prompting, context and guardrails alone — and good practices discovered in one sub-team never reached the others. We had the tools; we didn’t have a way to share what worked across all five stacks.
My role
Engineering Manager, COE UI — created LLMKIT and lead its design, rollout and roadmap, working closely with the sub-team leads on every stack.
Approach
- Shared AI context. An AI context generation system — Repomix extraction, a two-tier markdown architecture (project-level, then stack-specific), and one convention for where context lives, enforced across the monorepo.
- One kit, five stacks. Installs cleanly on every stack, respects each stack’s idioms, and still behaves as one coherent system — refined through repeated test-and-break cycles with the sub-team leads.
- Built for adoption. Teams notifications for installs, updates and releases, driven by an editable config, plus deliberate attention to day-to-day developer experience.
- Unified usage history. A layer that reads Claude Code, Codex and Copilot history natively, normalizes once at the end, and merges it into one provider-agnostic feed — powering a team dashboard. Legacy history is copied, never moved.
- Cross-platform parity. A macOS-vs-Windows audit, fixes, and direct verification on Windows 11, reported as an honest scorecard.
Where it stands
LLMKIT has become the thing tying the whole COE UI team’s AI workflow together:
- Consistent setup across all five stacks
- A shared AI-context convention, so every engineer’s assistant starts from real project context rather than guesswork
- Visibility built in, so the team knows what’s changing and when
- One view of AI-assisted development usage across three assistants
- A sixth environment, Windows 11, with its remaining gaps tracked openly
The part I’m proudest of isn’t the tooling itself — it’s that a good practice discovered in one sub-team can now reach every other sub-team, instead of staying siloed. LLMKIT is still in active development.
The build, in detail
The Building LLMKIT series covers each milestone: