Workloom

Every AI coding assistant an org uses — Claude Code, Codex, Copilot, Gemini, Cursor — has its own config format and file locations, and those configs get copy-pasted between repos with no versioning, no review, and no drift detection. As the brief put it: “the pre-npm era of AI configuration, getting worse every week.”

1. 01

   Modeled the system as “npm + Git objects + Terraform plan/apply + provider renderers + CI drift detection,” purpose-built for AI config.

2. 02

   Let teams author canonical packages of instructions and skills, then pin versions to repos through workspace manifests.

3. 03

   Rendered provider-specific files deterministically so each tool gets exactly the config it expects — semantics live in the renderer, not a dumb file sync.

4. 04

   Added sync to detect and resolve drift, governed by trust gates and audit logs for team and enterprise use.

Workloom deliberately avoided competing prototypes in favor of one design system that matured over time — captured in a generated design spec with a “Burnt Clay” primary in OKLCH and dark-mode canonicals. The git history reads as a sequence rather than parallel experiments: first a dual light/dark token system with a shadcn sidebar migrated across every route, then runtime-switchable brand palettes and a fixed top bar, and finally an anonymous public “front door” — an install-and-quickstart onboarding surface — layered on top of the authenticated app and hardened through automated design critique.